Release NVIDIA WakeWord Trainer v27

VERSION

@@ -1 +1 @@
-23
+27

WHATS_NEW.md

@@ -1,2 +1 @@
-- Fixed NVIDIA v22 training runs remaining stuck immediately after Start Session instead of launching the training worker.
-- Corrected the worker-state handoff for both manual and automatic training, with regression coverage for the complete startup path.
+- Added English accent emphasis for Mixed English, Australian, American, British, Canadian, Irish, Scottish, New Zealand, Indian, and South African voices. Qwen shapes the selected accent and MOSS carries it into cloned references, with the setting available in both manual and automatic training.

cli/shell.functions

@@ -12,6 +12,7 @@ DEFAULT_SAMPLES=50000
 DEFAULT_BATCH_SIZE=100
 DEFAULT_TRAINING_STEPS=40000
 DEFAULT_LANGUAGE=en
+DEFAULT_ENGLISH_ACCENT=mixed
 DEFAULT_TTS_MODE=hybrid
 DEFAULT_TTS_VOICE_COUNT=128
 
@@ -21,6 +22,7 @@ DEFAULT_TTS_VOICE_COUNT=128
 : "${BATCH_SIZE:=${DEFAULT_BATCH_SIZE}}"
 : "${TRAINING_STEPS:=${DEFAULT_TRAINING_STEPS}}"
 : "${LANGUAGE:=${DEFAULT_LANGUAGE}}"
+: "${ENGLISH_ACCENT:=${DEFAULT_ENGLISH_ACCENT}}"
 : "${TTS_MODE:=${DEFAULT_TTS_MODE}}"
 : "${TTS_VOICE_COUNT:=${DEFAULT_TTS_VOICE_COUNT}}"
 : "${CLEANUP_WORK_DIR:=false}"

cli/tts_generate_samples.py

@@ -16,6 +16,7 @@ import math
 import os
 import random
 import shutil
+import signal
 import subprocess
 import sys
 import wave
@@ -30,20 +31,24 @@ if str(ROOT_DIR) not in sys.path:
     sys.path.insert(0, str(ROOT_DIR))
 
 from tts_config import (  # noqa: E402
+    DEFAULT_ENGLISH_ACCENT,
     DEFAULT_TTS_MODE,
+    ENGLISH_ACCENTS,
     ENGINE_MOSS,
     ENGINE_OMNIVOICE,
     ENGINE_PIPER,
     ENGINE_QWEN3,
+    MIXED_ENGLISH_ACCENTS,
     QWEN_LANGUAGE_NAMES,
     distribute_samples,
     engines_for_language,
     language_for_engine,
+    normalize_english_accent,
     normalize_tts_mode,
 )
 
 
-GENERATOR_VERSION = "modern-tts-v15-four-provider-direct-corpus-safe-limits"
+GENERATOR_VERSION = "modern-tts-v17-four-provider-direct-corpus-safe-limits-english-accent-emphasis"
 VOICE_BANK_VERSION = "modern-tts-voice-bank-v1-native-random-qualified-single-utterance"
 COMPATIBLE_VOICE_BANK_VERSIONS = {
     VOICE_BANK_VERSION,
@@ -65,6 +70,8 @@ DIRECT_CANDIDATE_FACTORS = {
     ENGINE_MOSS: 1.25,
     ENGINE_PIPER: 1.05,
 }
+NORMALIZATION_TIMEOUT_SECONDS = 30.0
+NORMALIZATION_PROGRESS_INTERVAL = 100
 
 CARRIER_PROMPT_TEMPLATES = {
     "ar": "بصوت هادئ وطبيعي أقول {phrase} بوضوح، ثم أواصل الحديث بإيقاع ثابت.",
@@ -122,6 +129,38 @@ def run_with_batch_retry(
         run(retry_command, env=env)
 
 
+def run_normalization_ffmpeg(command: list[str], timeout: float) -> int | None:
+    """Run one conversion without allowing a stuck file read to block training."""
+
+    process = subprocess.Popen(
+        command,
+        stdin=subprocess.DEVNULL,
+        start_new_session=True,
+    )
+    try:
+        return process.wait(timeout=timeout)
+    except subprocess.TimeoutExpired:
+        # Do not use subprocess.run(timeout=...) here. On POSIX it performs an
+        # unbounded wait after killing the child, which can still freeze the
+        # trainer when a process is stuck in filesystem I/O.
+        try:
+            os.killpg(process.pid, signal.SIGKILL)
+        except ProcessLookupError:
+            pass
+        except OSError:
+            try:
+                process.kill()
+            except ProcessLookupError:
+                pass
+        try:
+            process.wait(timeout=2.0)
+        except subprocess.TimeoutExpired:
+            # The process may remain in uninterruptible I/O until the kernel
+            # releases it. The next candidate can still be processed safely.
+            pass
+        return None
+
+
 def write_jsonl(path: Path, entries: list[dict]) -> None:
     path.parent.mkdir(parents=True, exist_ok=True)
     with path.open("w", encoding="utf-8") as stream:
@@ -164,7 +203,11 @@ def stable_prompt_text(phrase: str, language: str = "en") -> str:
     return clean + "."
 
 
-def qwen_descriptions(language_name: str, count: int) -> list[str]:
+def qwen_descriptions(
+    language_name: str,
+    count: int,
+    english_accent: str = DEFAULT_ENGLISH_ACCENT,
+) -> list[str]:
     genders = ("female", "male")
     ages = ("child", "teenager", "young adult", "middle-aged adult", "elderly adult")
     pitches = ("low pitch", "medium pitch", "high pitch")
@@ -180,16 +223,30 @@ def qwen_descriptions(language_name: str, count: int) -> list[str]:
     weights = ("light", "balanced", "compact", "full-bodied", "resonant")
     combinations = list(product(genders, ages, pitches, deliveries, textures, paces, weights))
     descriptions = []
+    accent_cycle: tuple[str, ...] = ()
+    if language_name == "English":
+        selected_accent = normalize_english_accent(english_accent, "en")
+        accent_cycle = (
+            MIXED_ENGLISH_ACCENTS
+            if selected_accent == DEFAULT_ENGLISH_ACCENT
+            else (selected_accent,)
+        )
     # Walking the Cartesian product sequentially clusters the leading traits
     # (the first 375 combinations are all female).  A coprime stride retains a
     # deterministic, non-repeating order while balancing every trait early.
     for index in range(count):
         combination_index = (index * VOICE_PROFILE_STRIDE) % len(combinations)
         gender, age, pitch, delivery, texture, pace, weight = combinations[combination_index]
+        language_style = f"native {language_name}"
+        if accent_cycle:
+            selected_accent = accent_cycle[index % len(accent_cycle)]
+            language_style = (
+                f"English with a natural {ENGLISH_ACCENTS[selected_accent]} accent"
+            )
         descriptions.append(
             f"A distinct {age} {gender} speaker with a {texture} timbre, "
-            f"{pitch}, {weight} vocal weight, and {delivery}, speaking native "
-            f"{language_name} at a {pace} pace. Say only the supplied text once."
+            f"{pitch}, {weight} vocal weight, and {delivery}, speaking "
+            f"{language_style} at a {pace} pace. Say only the supplied text once."
         )
     return descriptions
 
@@ -248,6 +305,10 @@ def valid_sample(path: Path) -> bool:
 class Generator:
     def __init__(self, args: argparse.Namespace):
         self.args = args
+        self.english_accent = normalize_english_accent(
+            getattr(args, "english_accent", DEFAULT_ENGLISH_ACCENT),
+            args.language,
+        )
         self.spoken_phrase = args.phrase.replace("_", " ").strip()
         self.data_dir = args.data_dir.resolve()
         self.output_dir = args.output_dir.resolve()
@@ -304,6 +365,7 @@ class Generator:
             "generator_version": GENERATOR_VERSION,
             "phrase": self.args.phrase,
             "language": self.args.language,
+            "english_accent": self.english_accent,
             "tts_mode": self.args.tts_mode,
             "samples": self.args.samples,
             "engines": engines,
@@ -1029,7 +1091,11 @@ class Generator:
         self.direct_attempt[engine] += count
         rng = random.Random(24051984 + start + sum(ord(ch) for ch in engine + prefix))
         descriptions = (
-            qwen_descriptions(QWEN_LANGUAGE_NAMES[self.args.language], start + count)[start:]
+            qwen_descriptions(
+                QWEN_LANGUAGE_NAMES[self.args.language],
+                start + count,
+                self.english_accent,
+            )[start:]
             if engine == ENGINE_QWEN3
             else []
         )
@@ -1247,7 +1313,9 @@ class Generator:
     def normalize(self, paths: list[Path], start_index: int, limit: int) -> list[Path]:
         accepted = []
         self.final_dir.mkdir(parents=True, exist_ok=True)
-        for path in paths:
+        candidate_count = len(paths)
+        log(f"→ Normalizing up to {limit} accepted clip(s) from {candidate_count} candidate(s)")
+        for processed, path in enumerate(paths, start=1):
             if len(accepted) >= limit:
                 break
             final_path = self.final_dir / f"{start_index + len(accepted)}.wav"
@@ -1258,6 +1326,7 @@ class Generator:
                 "-hide_banner",
                 "-loglevel",
                 "error",
+                "-nostdin",
                 "-y",
                 "-i",
                 str(path),
@@ -1272,18 +1341,39 @@ class Generator:
                 "pcm_s16le",
                 str(temp_path),
             ]
-            try:
-                subprocess.run(command, check=True)
-            except subprocess.CalledProcessError:
+            return_code = run_normalization_ffmpeg(
+                command,
+                timeout=NORMALIZATION_TIMEOUT_SECONDS,
+            )
+            converted = return_code == 0
+            if return_code is None:
                 temp_path.unlink(missing_ok=True)
-                continue
-            digest = hashlib.sha256(temp_path.read_bytes()).hexdigest() if temp_path.is_file() else ""
-            if valid_sample(temp_path) and digest and digest not in self.accepted_hashes:
-                temp_path.replace(final_path)
-                self.accepted_hashes.add(digest)
-                accepted.append(final_path)
-            else:
+                log(
+                    f"⚠️ Normalization timed out after "
+                    f"{NORMALIZATION_TIMEOUT_SECONDS:g}s; skipping {path.name}"
+                )
+            elif return_code != 0:
                 temp_path.unlink(missing_ok=True)
+                log(f"⚠️ ffmpeg rejected {path.name} (exit {return_code}); skipping it")
+
+            if converted:
+                digest = hashlib.sha256(temp_path.read_bytes()).hexdigest() if temp_path.is_file() else ""
+                if valid_sample(temp_path) and digest and digest not in self.accepted_hashes:
+                    temp_path.replace(final_path)
+                    self.accepted_hashes.add(digest)
+                    accepted.append(final_path)
+                else:
+                    temp_path.unlink(missing_ok=True)
+
+            if (
+                processed % NORMALIZATION_PROGRESS_INTERVAL == 0
+                or processed == candidate_count
+                or len(accepted) >= limit
+            ):
+                log(
+                    f"Normalization progress: {len(accepted)}/{limit} accepted "
+                    f"({processed}/{candidate_count} candidate(s) checked)"
+                )
         return accepted
 
     def generate(self) -> None:
@@ -1301,6 +1391,8 @@ class Generator:
         self.final_dir.mkdir(parents=True, exist_ok=True)
         plan = distribute_samples(self.args.samples, engines)
         log(f"===== Direct TTS corpus plan ({self.args.tts_mode}, {self.args.language}) =====")
+        if self.args.language == "en" and ENGINE_QWEN3 in plan:
+            log(f"   English accent emphasis: {self.english_accent}")
         for engine, count in plan.items():
             log(f"   {engine}: {count} sample(s)")
         log(
@@ -1388,6 +1480,7 @@ class Generator:
                 "reusable_profile_bank": False,
                 "moss_unique_accepted_carriers": True,
                 "piper_all_model_speakers": True,
+                "english_accent_emphasis": self.english_accent,
             },
             "qa": {
                 "audio_format": "16 kHz mono PCM16 WAV",
@@ -1416,6 +1509,10 @@ def parser() -> argparse.ArgumentParser:
     result = argparse.ArgumentParser()
     result.add_argument("phrase")
     result.add_argument("--language", default="en")
+    result.add_argument(
+        "--english-accent",
+        default=os.environ.get("MWW_ENGLISH_ACCENT", DEFAULT_ENGLISH_ACCENT),
+    )
     result.add_argument("--tts-mode", default=DEFAULT_TTS_MODE)
     result.add_argument("--samples", type=int, default=50000)
     result.add_argument("--batch-size", type=int, default=8)
@@ -1435,6 +1532,7 @@ def parser() -> argparse.ArgumentParser:
 def main() -> int:
     args = parser().parse_args()
     args.language = args.language.strip().lower().replace("-", "_")
+    args.english_accent = normalize_english_accent(args.english_accent, args.language)
     args.tts_mode = normalize_tts_mode(args.tts_mode)
     if args.samples < 1:
         raise SystemExit("--samples must be positive")

cli/tts_moss_worker.py

@@ -69,7 +69,6 @@ def main() -> int:
             text=str(item["text"]),
             output_audio_path=str(output_path),
             mode="voice_clone",
-            prompt_text=str(item["ref_text"]),
             prompt_audio_path=str(item["ref_audio"]),
             reference_audio_path=None,
             text_tokenizer_path=None,

cli/wake_word_sample_generator

@@ -4,7 +4,7 @@ set -euo pipefail
 PROGPATH="$(realpath "$0")"
 PROGDIR="$(dirname "${PROGPATH}")"
 
-KNOWN_ARGS=( samples batch-size data-dir language tts-mode tts-voice-count )
+KNOWN_ARGS=( samples batch-size data-dir language english-accent tts-mode tts-voice-count )
 # shellcheck source=/dev/null
 source "${PROGDIR}/shell.functions"
 WAKE_WORD="${POSITIONAL_ARGS[0]:-}"
@@ -17,12 +17,14 @@ fi
 if [ "${HELP}" == "true" ] || [ -z "${WAKE_WORD}" ] ; then
     cat <<EOF >&2
 Usage: $0 [ --samples=<samples> ] [ --batch-size=<batch_size> ]
-          [ --language=<lang> ] [ --tts-mode=<modern|hybrid|piper> ]
+          [ --language=<lang> ] [ --english-accent=<accent> ]
+          [ --tts-mode=<modern|hybrid|piper> ]
           [ --tts-voice-count=<voices> ] <wake_word>
 
 --samples:            Number of samples to generate. Default: ${DEFAULT_SAMPLES}
 --batch-size:         Generation batch size. Default: ${DEFAULT_BATCH_SIZE}
 --language:           TTS language code. Default: ${DEFAULT_LANGUAGE}
+--english-accent:     English accent emphasis. Default: ${DEFAULT_ENGLISH_ACCENT}
 --tts-mode:           modern, hybrid, or piper. Default: ${DEFAULT_TTS_MODE}
 --tts-voice-count:    Deprecated compatibility option; direct generation ignores it.
 <wake_word>           Required phrase to synthesize.
@@ -38,17 +40,31 @@ case "${TTS_MODE}" in
         ;;
 esac
 
+LANGUAGE="$(echo "${LANGUAGE}" | tr '[:upper:]' '[:lower:]')"
+ENGLISH_ACCENT="$(echo "${ENGLISH_ACCENT}" | tr '[:upper:] -' '[:lower:]__')"
+if [ "${LANGUAGE}" != "en" ]; then
+    ENGLISH_ACCENT="mixed"
+fi
+case "${ENGLISH_ACCENT}" in
+    mixed|australian|american|british|canadian|irish|scottish|new_zealand|indian|south_african) ;;
+    *)
+        echo "ERROR: unsupported --english-accent '${ENGLISH_ACCENT}'." >&2
+        exit 2
+        ;;
+esac
+
 WORK_DIR="${DATA_DIR}/work"
 SAMPLES_DIR="${WORK_DIR}/wake_word_samples"
 mkdir -p "${WORK_DIR}"
 
 START_TS=$EPOCHSECONDS
-echo "===== Generating ${SAMPLES} wake-word samples (language=${LANGUAGE}, tts=${TTS_MODE}) ====="
+echo "===== Generating ${SAMPLES} wake-word samples (language=${LANGUAGE}, accent=${ENGLISH_ACCENT}, tts=${TTS_MODE}) ====="
 
 python3 "${PROGDIR}/tts_generate_samples.py" "${WAKE_WORD}" \
     --samples="${SAMPLES}" \
     --batch-size="${BATCH_SIZE}" \
     --language="${LANGUAGE}" \
+    --english-accent="${ENGLISH_ACCENT}" \
     --tts-mode="${TTS_MODE}" \
     --voice-count="${TTS_VOICE_COUNT}" \
     --data-dir="${DATA_DIR}" \

frontend/src/TrainerApp.vue

@@ -70,7 +70,7 @@ const dataCategories = computed(() => {
   return Array.from(groups, ([name, items]) => ({ name, items }));
 });
 
-watch(() => trainer.language, ensureSupportedTtsMode);
+watch([() => trainer.language, () => trainer.ttsMode], ensureSupportedTtsMode);
 watch(() => trainer.toast.serial, () => window.setTimeout(() => { trainer.toast.message = ""; }, 4500));
 watch(consoleLines, async () => {
   if (!consoleFollowing.value) return;
@@ -162,6 +162,7 @@ function sampleSubtitle(item: AudioItem): string {
   return rows.join(" · ") || "Training sample";
 }
 function wordJsonUrl(item: JsonRecord): string { return String(item.json_url || item.url || item.jsonUrl || ""); }
+function wordEsphomeJsonUrl(item: JsonRecord): string { return String(item.esphome_json_url || item.esphomeJsonUrl || ""); }
 function wordModelUrl(item: JsonRecord): string { return String(item.model_url || item.modelUrl || ""); }
 function consoleTone(line: string): string {
   const value = line.trim().toLowerCase();
@@ -196,6 +197,7 @@ function consoleTone(line: string): string {
             <div class="form-grid phrase-form">
               <label class="field wide"><span>Wake phrase</span><input v-model="trainer.phrase" type="text" placeholder='e.g. "hey tater"' :disabled="Boolean(trainer.session.safe_word) || isBusy('session')" @keyup.enter="startSession" /></label>
               <label class="field"><span>Language</span><select v-model="trainer.language" :disabled="Boolean(trainer.session.safe_word) || isBusy('session')"><option v-for="item in trainer.languages" :key="item.code" :value="item.code">{{ item.label }}</option></select><small>{{ ttsRoute }}</small></label>
+              <label v-if="trainer.language === 'en' && trainer.ttsMode !== 'piper'" class="field"><span>English accent emphasis</span><select v-model="trainer.englishAccent" :disabled="Boolean(trainer.session.safe_word) || isBusy('session')"><option v-for="accent in trainer.englishAccents" :key="accent.code" :value="accent.code">{{ accent.label }}</option></select><small>Qwen emphasizes this accent; MOSS carries it through accepted references. OmniVoice and Piper keep broad English coverage.</small></label>
               <label class="field"><span>TTS source</span><select v-model="trainer.ttsMode" :disabled="Boolean(trainer.session.safe_word) || isBusy('session')">
                 <option value="hybrid" :disabled="!trainer.languages.find((item) => item.code === trainer.language)?.engines?.includes('piper')">Four-provider ensemble · recommended</option>
                 <option value="modern" :disabled="!trainer.languages.find((item) => item.code === trainer.language)?.engines?.some((engine) => engine !== 'piper')">Modern only · no Piper</option>
@@ -223,6 +225,7 @@ function consoleTone(line: string): string {
             <div class="form-grid">
               <label class="field"><span>Wake phrase</span><input v-model="trainer.autoForm.wake_phrase" type="text" /></label>
               <label class="field"><span>STT language</span><input v-model="trainer.autoForm.language" type="text" /></label>
+              <label v-if="String(trainer.autoForm.language).toLowerCase().startsWith('en')" class="field"><span>English accent emphasis</span><select v-model="trainer.autoForm.english_accent"><option v-for="accent in trainer.englishAccents" :key="accent.code" :value="accent.code">{{ accent.label }}</option></select><small>Used when Auto Training needs to regenerate English TTS.</small></label>
               <label class="field wide"><span>STT engine</span><select v-model="trainer.autoForm.stt_engine"><option v-for="engine in sttEngines" :key="engine.id || engine.value" :value="engine.id || engine.value">{{ engine.label || engine.name || engine.id }}</option></select><small>{{ sttEngines.find((row) => (row.id || row.value) === trainer.autoForm.stt_engine)?.description || "Runs locally on this trainer." }}</small></label>
               <label class="field"><span>Minimum transcript characters</span><input v-model.number="trainer.autoForm.minimum_transcript_chars" min="1" max="100" type="number" /></label>
             </div>
@@ -265,6 +268,7 @@ function consoleTone(line: string): string {
             <div v-if="!selectedSamples.length" class="empty-state">No {{ trainer.sampleBucket }} samples saved yet.</div>
             <div v-else class="audio-list compact-list"><article v-for="item in pagedSamples" :key="item.saved_as" class="audio-card">
               <header><div><strong>{{ item.saved_as }}</strong><small>{{ sampleSubtitle(item) }}</small></div><div class="row"><span v-if="item.trimmed" class="pill warning">Trimmed</span><span class="pill" :class="trainer.sampleBucket === 'personal' ? 'success' : 'error'">{{ trainer.sampleBucket === "personal" ? "Positive" : "Negative" }}</span></div></header>
+              <div v-if="item.transcript" class="transcript"><b>STT</b> {{ item.transcript }}</div><div v-if="item.auto_review_guided_transcript" class="transcript"><b>Guided wake check</b> {{ item.auto_review_guided_transcript }}</div>
               <audio controls preload="none" :src="itemAudioUrl(item, trainer.sampleBucket)" />
               <footer><span>{{ describeFormat(item.final_format) }}</span><div><button type="button" @click="openTrim(item, trainer.sampleBucket)">Trim</button><button v-if="item.trimmed" type="button" @click="revertSample(item, trainer.sampleBucket)">Revert</button><button type="button" class="button danger ghost" :disabled="isBusy('review')" @click="removeSample(item, trainer.sampleBucket)">Remove</button></div></footer>
             </article></div>
@@ -301,6 +305,11 @@ function consoleTone(line: string): string {
             <div v-if="!trainer.wakeWords.length" class="empty-state">Train a wake word and its package will appear here.</div>
             <div v-else class="word-list"><article v-for="word in trainer.wakeWords" :key="word.key || wordJsonUrl(word)"><div><strong>{{ word.label || word.name || "Trained wake word" }}</strong><a v-if="wordJsonUrl(word)" :href="wordJsonUrl(word)" target="_blank" rel="noreferrer">JSON · {{ wordJsonUrl(word) }}</a><span v-else class="muted">JSON package URL unavailable</span><a v-if="wordModelUrl(word)" :href="wordModelUrl(word)" target="_blank" rel="noreferrer">Model · {{ wordModelUrl(word) }}</a><div class="meta-row"><span v-if="word.language">{{ word.language }}</span><span v-if="word.trained_at">{{ formatTimestamp(word.trained_at) }}</span><span v-if="word.recall !== undefined">recall {{ word.recall }}</span></div></div><button type="button" :disabled="!wordJsonUrl(word)" @click="copyWakeWord(wordJsonUrl(word))">Copy URL</button></article></div>
           </section>
+          <div class="native-notice esphome-notice"><strong>ESPHome</strong><span>Strict micro_wake_word manifest without Tater Native or calibration extensions.</span></div>
+          <section class="panel compatibility-panel"><header class="panel-head"><div class="number">ESP</div><div><h3>ESPHome JSON</h3><p>Use this URL as the model in an ESPHome micro_wake_word configuration.</p></div></header>
+            <div v-if="!trainer.wakeWords.length" class="empty-state">ESPHome links appear after a wake word is trained.</div>
+            <div v-else class="word-list"><article v-for="word in trainer.wakeWords" :key="`esphome-${word.key || wordEsphomeJsonUrl(word)}`"><div><strong>{{ word.label || word.name || "Trained wake word" }}</strong><a v-if="wordEsphomeJsonUrl(word)" :href="wordEsphomeJsonUrl(word)" target="_blank" rel="noreferrer">ESPHome JSON · {{ wordEsphomeJsonUrl(word) }}</a><span v-else class="muted">ESPHome package URL unavailable</span><div class="meta-row"><span>Schema v2</span><span>Same TFLite model</span></div></div><button type="button" :disabled="!wordEsphomeJsonUrl(word)" @click="copyWakeWord(wordEsphomeJsonUrl(word))">Copy ESPHome URL</button></article></div>
+          </section>
         </template>
       </template>
     </main>

frontend/src/trainer.css

@@ -147,6 +147,9 @@ button:disabled { opacity: .43; cursor: not-allowed; }
 .progress-track i { display: block; height: 100%; border-radius: inherit; background: linear-gradient(90deg, var(--orange), var(--violet)); transition: width .2s ease; }
 .native-notice { display: flex; align-items: center; gap: 12px; padding: 14px 18px; color: var(--muted); font-size: 12px; }
 .native-notice strong { color: var(--green); }
+.esphome-notice strong { color: var(--orange-2); }
+.compatibility-panel { padding-top: 19px; }
+.compatibility-panel .panel-head { margin-bottom: 15px; }
 .word-list article { display: flex; justify-content: space-between; align-items: center; gap: 20px; padding: 16px; border: 1px solid var(--line); border-radius: 15px; background: rgba(18,18,19,.64); }
 .word-list article > div { display: grid; min-width: 0; gap: 6px; }
 .word-list a { overflow-wrap: anywhere; color: var(--orange-2); font-size: 11px; text-decoration: none; }

frontend/src/trainerStore.ts

@@ -1,6 +1,7 @@
 import { computed, reactive } from "vue";
 import { getJson, postJson, putJson, request, type JsonRecord } from "./api";
 import type {
+  AccentOption,
   AudioItem,
   AutoTrainForm,
   AutoTrainPayload,
@@ -26,6 +27,7 @@ const defaultAutoForm = (): AutoTrainForm => ({
   enabled: false,
   wake_phrase: "",
   language: "en",
+  english_accent: "mixed",
   stt_engine: "faster_whisper",
   minimum_transcript_chars: 2,
   delete_confirmed_wakes: false,
@@ -43,8 +45,21 @@ export const trainer = reactive({
   busy: new Set<string>(),
   phrase: "",
   language: "en",
+  englishAccent: "mixed",
   ttsMode: "hybrid",
   languages: [{ code: "en", label: "English (en)", engines: ["omnivoice"] }] as LanguageOption[],
+  englishAccents: [
+    { code: "mixed", label: "Mixed English" },
+    { code: "australian", label: "Australian" },
+    { code: "american", label: "American" },
+    { code: "british", label: "British" },
+    { code: "canadian", label: "Canadian" },
+    { code: "irish", label: "Irish" },
+    { code: "scottish", label: "Scottish" },
+    { code: "new_zealand", label: "New Zealand" },
+    { code: "indian", label: "Indian" },
+    { code: "south_african", label: "South African" },
+  ] as AccentOption[],
   session: {} as SessionPayload,
   samples: emptySamples(),
   captured: emptyCaptured(),
@@ -123,8 +138,12 @@ function applySession(payload: SessionPayload): void {
   if (Array.isArray(payload.available_languages) && payload.available_languages.length) {
     trainer.languages = payload.available_languages;
   }
+  if (Array.isArray(payload.available_english_accents) && payload.available_english_accents.length) {
+    trainer.englishAccents = payload.available_english_accents;
+  }
   if (payload.raw_phrase) trainer.phrase = payload.raw_phrase;
   if (payload.language) trainer.language = payload.language;
+  if (payload.english_accent) trainer.englishAccent = payload.english_accent;
   if (payload.tts_mode) trainer.ttsMode = payload.tts_mode;
   if (payload.training) trainer.training = payload.training;
 }
@@ -145,6 +164,7 @@ export async function startSession(): Promise<void> {
     const payload = await postJson<SessionPayload>("/api/start_session", {
       phrase: trainer.phrase.trim(),
       language: trainer.language,
+      english_accent: trainer.englishAccent,
       tts_mode: trainer.ttsMode,
     });
     applySession(payload);
@@ -193,6 +213,7 @@ export function ensureSupportedTtsMode(): void {
   if (trainer.ttsMode === "modern" && !modern) trainer.ttsMode = "piper";
   if (trainer.ttsMode === "hybrid" && !(modern && piper)) trainer.ttsMode = modern ? "modern" : "piper";
   if (trainer.ttsMode === "piper" && !piper) trainer.ttsMode = "modern";
+  if (trainer.language !== "en" || trainer.ttsMode === "piper") trainer.englishAccent = "mixed";
 }
 
 export async function refreshSamples(quiet = false): Promise<SamplesPayload> {
@@ -340,6 +361,8 @@ function applyAuto(payload: AutoTrainPayload, populate: boolean): void {
   trainer.autoForm = { ...defaultAutoForm(), ...(payload.config || {}) };
   if (!trainer.autoForm.wake_phrase) trainer.autoForm.wake_phrase = trainer.session.raw_phrase || "";
   if (!trainer.autoForm.language) trainer.autoForm.language = trainer.session.language || "en";
+  if (!trainer.autoForm.english_accent) trainer.autoForm.english_accent = trainer.session.english_accent || "mixed";
+  if (!String(trainer.autoForm.language).toLowerCase().startsWith("en")) trainer.autoForm.english_accent = "mixed";
 }
 
 export async function refreshAuto(populate = false): Promise<AutoTrainPayload> {

frontend/src/types.ts

@@ -10,6 +10,11 @@ export interface LanguageOption extends JsonRecord {
   quality?: string;
 }
 
+export interface AccentOption extends JsonRecord {
+  code: string;
+  label: string;
+}
+
 export interface TrainingState extends JsonRecord {
   running: boolean;
   exit_code: number | null;
@@ -20,9 +25,11 @@ export interface SessionPayload extends JsonRecord {
   safe_word?: string;
   raw_phrase?: string;
   language?: string;
+  english_accent?: string;
   tts_mode?: string;
   takes_received?: number;
   available_languages?: LanguageOption[];
+  available_english_accents?: AccentOption[];
   training?: TrainingState;
 }
 
@@ -31,6 +38,11 @@ export interface AudioItem extends JsonRecord {
   original_name?: string;
   audio_url?: string;
   final_format?: JsonRecord;
+  transcript?: string;
+  transcribed_at?: string;
+  auto_review_guided_transcript?: string;
+  auto_review_stt_engine?: string;
+  auto_review_stt_model?: string;
 }
 
 export interface SamplesPayload extends JsonRecord {
@@ -51,6 +63,7 @@ export interface AutoTrainForm extends JsonRecord {
   enabled: boolean;
   wake_phrase: string;
   language: string;
+  english_accent: string;
   stt_engine: string;
   minimum_transcript_chars: number;
   delete_confirmed_wakes: boolean;
@@ -76,6 +89,8 @@ export interface WakeWordItem extends JsonRecord {
   url?: string;
   json_url?: string;
   jsonUrl?: string;
+  esphome_json_url?: string;
+  esphomeJsonUrl?: string;
   model_url?: string;
   modelUrl?: string;
 }

tests/test_auto_train.py

@@ -315,6 +315,10 @@ class AutoTrainTests(unittest.TestCase):
         self.assertEqual(metadata["transcript"], "turn on the kitchen lights")
         self.assertEqual(metadata["auto_review_stt_engine"], "faster_whisper")
         self.assertEqual(metadata["auto_review_stt_model"], "small.en")
+        sample_item = trainer._sample_item_from_path(negatives[0], "negative")
+        self.assertEqual(sample_item["transcript"], "turn on the kitchen lights")
+        self.assertEqual(sample_item["auto_review_stt_engine"], "faster_whisper")
+        self.assertEqual(sample_item["auto_review_stt_model"], "small.en")
         self.assertEqual(trainer.AUTO_TRAIN_STATE["pending_negative_count"], 1)
 
     def test_matching_phrase_stays_in_manual_review_inbox(self):
@@ -467,9 +471,30 @@ class AutoTrainTests(unittest.TestCase):
         self.assertTrue(metadata["auto_positive"])
         self.assertEqual(metadata["review_status"], "auto_approved_personal")
         self.assertEqual(metadata["transcript"], "hey tater")
+        sample_item = trainer._sample_item_from_path(positives[0], "personal")
+        self.assertEqual(sample_item["transcript"], "hey tater")
         self.assertFalse(list(trainer.NEGATIVE_DIR.glob("*.wav")))
         self.assertEqual(trainer.AUTO_TRAIN_STATE["pending_negative_count"], 0)
 
+    def test_guided_stt_remains_visible_after_positive_auto_sort(self):
+        self.add_capture(event_type="close_miss")
+        trainer.AUTO_TRAIN_CONFIG["promote_close_misses"] = True
+        with (
+            patch.object(trainer, "_transcribe_capture", return_value="Hey, haters."),
+            patch.object(
+                trainer,
+                "_transcribe_capture_with_faster_whisper_guided",
+                return_value="Hey Tater",
+            ),
+        ):
+            trainer._auto_review_capture("wake.wav")
+
+        positives = list(trainer.PERSONAL_DIR.glob("*.wav"))
+        self.assertEqual(len(positives), 1)
+        sample_item = trainer._sample_item_from_path(positives[0], "personal")
+        self.assertEqual(sample_item["transcript"], "Hey, haters.")
+        self.assertEqual(sample_item["auto_review_guided_transcript"], "Hey Tater")
+
     def test_close_miss_without_phrase_stays_in_inbox(self):
         audio_path = self.add_capture(event_type="close_miss")
         trainer.AUTO_TRAIN_CONFIG["promote_close_misses"] = True
@@ -583,6 +608,56 @@ class AutoTrainTests(unittest.TestCase):
         self.assertEqual(len(rows), 1)
         self.assertEqual(rows[0]["url"], rows[0]["json_url"])
         self.assertTrue(rows[0]["json_url"].endswith("/api/trained_wake_words/hey_tater.json"))
+        self.assertTrue(
+            rows[0]["esphome_json_url"].endswith(
+                "/api/trained_wake_words/hey_tater.esphome.json"
+            )
+        )
+
+    def test_esphome_manifest_route_removes_tater_extensions(self):
+        with tempfile.TemporaryDirectory() as directory:
+            trained_dir = Path(directory)
+            (trained_dir / "hey_tater.tflite").write_bytes(b"model")
+            metadata = {
+                "type": "micro",
+                "wake_word": "hey tater",
+                "label": "Hey Tater",
+                "author": "Tater Totterson",
+                "website": "https://example.com",
+                "model": "hey_tater.tflite",
+                "trained_languages": ["en"],
+                "version": 2,
+                "model_format": "tflite_stream_state_internal_quant",
+                "quantization": "int8",
+                "sample_rate": 16000,
+                "micro": {
+                    "probability_cutoff": 0.97,
+                    "sliding_window_size": 5,
+                    "feature_step_size": 10,
+                    "tensor_arena_size": 30000,
+                    "minimum_esphome_version": "2024.7.0",
+                },
+                "tater_native": {"format_version": 1},
+                "calibration": {"recall": 0.99},
+            }
+            (trained_dir / "hey_tater.json").write_text(
+                json.dumps(metadata),
+                encoding="utf-8",
+            )
+            with (
+                patch.object(trainer, "TRAINED_WAKE_WORDS_DIR", trained_dir),
+                patch.object(trainer, "_sync_trained_wake_word_artifacts"),
+            ):
+                response = trainer.trained_wake_word_artifact(
+                    "hey_tater.esphome.json"
+                )
+
+        payload = json.loads(response.body)
+        self.assertEqual(set(payload), set(trainer.ESPHOME_MANIFEST_KEYS))
+        self.assertEqual(payload["micro"], metadata["micro"])
+        self.assertNotIn("label", payload)
+        self.assertNotIn("tater_native", payload)
+        self.assertNotIn("calibration", payload)
 
     def test_tater_notification_fails_when_trained_word_is_missing(self):
         trainer.AUTO_TRAIN_CONFIG["tater_link_token"] = "secret-token"

tests/test_modern_tts.py

@@ -1,10 +1,12 @@
 from __future__ import annotations
 
 import argparse
+import ast
 import importlib.util
 import json
 import math
 import shutil
+import signal
 import subprocess
 import tempfile
 import unittest
@@ -69,6 +71,44 @@ class ModernTtsTests(unittest.TestCase):
             ["--position_temperature", "5.0", "--class_temperature", "0.0"],
         )
 
+    def test_qwen_accent_emphasis_supports_specific_and_mixed_english(self) -> None:
+        australian = generator_module.qwen_descriptions("English", 4, "australian")
+        self.assertTrue(all("natural Australian accent" in row for row in australian))
+
+        mixed = generator_module.qwen_descriptions("English", 9, "mixed")
+        for label in (
+            "Australian",
+            "American",
+            "British",
+            "Canadian",
+            "Irish",
+            "Scottish",
+            "New Zealand",
+            "Indian",
+            "South African",
+        ):
+            self.assertTrue(any(f"natural {label} accent" in row for row in mixed))
+
+        german = generator_module.qwen_descriptions("German", 1, "australian")
+        self.assertIn("speaking native German", german[0])
+        self.assertNotIn("accent", german[0])
+
+    def test_moss_voice_clone_uses_audio_without_disallowed_prompt_text(self) -> None:
+        worker_path = REPO_ROOT / "cli" / "tts_moss_worker.py"
+        tree = ast.parse(worker_path.read_text(encoding="utf-8"))
+        inference_calls = [
+            node
+            for node in ast.walk(tree)
+            if isinstance(node, ast.Call)
+            and isinstance(node.func, ast.Attribute)
+            and node.func.attr == "inference"
+        ]
+
+        self.assertEqual(len(inference_calls), 1)
+        keywords = {keyword.arg for keyword in inference_calls[0].keywords}
+        self.assertIn("prompt_audio_path", keywords)
+        self.assertNotIn("prompt_text", keywords)
+
     def test_omnivoice_uses_a_hidden_stable_prompt_before_short_clone(self) -> None:
         with tempfile.TemporaryDirectory() as temp_dir:
             data_dir = Path(temp_dir)
@@ -586,6 +626,87 @@ class ModernTtsTests(unittest.TestCase):
             self.assertTrue((output_dir / ".generation_manifest.json").is_file())
             self.assertTrue(instance.cache_hit())
 
+    def test_normalization_times_out_bad_clip_and_continues(self) -> None:
+        with tempfile.TemporaryDirectory() as temp_dir:
+            data_dir = Path(temp_dir)
+            output_dir = data_dir / "work" / "wake_word_samples"
+            args = argparse.Namespace(
+                phrase="hey tater",
+                language="en",
+                tts_mode="modern",
+                samples=2,
+                batch_size=1,
+                voice_count=2,
+                data_dir=data_dir,
+                output_dir=output_dir,
+                ffmpeg="ffmpeg",
+                dry_run=False,
+            )
+            instance = generator_module.Generator(args)
+            raw_dir = instance.raw_dir / "qwen3"
+            raw_dir.mkdir(parents=True)
+            paths = [raw_dir / "bad.wav", raw_dir / "good.wav"]
+            for path in paths:
+                write_tone(path)
+                instance.speed_by_path[path.resolve()] = 1.0
+
+            calls = []
+
+            class FakeProcess:
+                def __init__(self, *, pid, timed_out):
+                    self.pid = pid
+                    self.timed_out = timed_out
+                    self.wait_calls = []
+
+                def wait(self, timeout=None):
+                    self.wait_calls.append(timeout)
+                    if self.timed_out and len(self.wait_calls) == 1:
+                        raise subprocess.TimeoutExpired(
+                            calls[0][0],
+                            generator_module.NORMALIZATION_TIMEOUT_SECONDS,
+                        )
+                    return -signal.SIGKILL if self.timed_out else 0
+
+                def kill(self):
+                    return None
+
+            processes = []
+
+            def fake_ffmpeg(command, **kwargs):
+                calls.append((command, kwargs))
+                temp_path = Path(command[-1])
+                if len(calls) == 1:
+                    temp_path.touch()
+                    process = FakeProcess(pid=12345, timed_out=True)
+                else:
+                    write_tone(temp_path)
+                    process = FakeProcess(pid=12346, timed_out=False)
+                processes.append(process)
+                return process
+
+            messages = []
+            with (
+                patch.object(generator_module.subprocess, "Popen", side_effect=fake_ffmpeg),
+                patch.object(generator_module.os, "killpg") as killpg,
+                patch.object(generator_module, "log", side_effect=messages.append),
+            ):
+                accepted = instance.normalize(paths, 0, 2)
+
+            self.assertEqual(len(calls), 2)
+            self.assertEqual(len(accepted), 1)
+            self.assertTrue((instance.final_dir / "0.wav").is_file())
+            self.assertFalse((instance.final_dir / "0.tmp.wav").exists())
+            self.assertIn("-nostdin", calls[0][0])
+            self.assertIs(calls[0][1]["stdin"], subprocess.DEVNULL)
+            self.assertTrue(calls[0][1]["start_new_session"])
+            self.assertEqual(
+                processes[0].wait_calls,
+                [generator_module.NORMALIZATION_TIMEOUT_SECONDS, 2.0],
+            )
+            killpg.assert_called_once_with(12345, signal.SIGKILL)
+            self.assertTrue(any("timed out" in message for message in messages))
+            self.assertTrue(any("1/2 accepted" in message for message in messages))
+
     def test_docker_and_ui_are_wired_for_modern_tts(self) -> None:
         for dockerfile in ("dockerfile", "dockerfile.blackwell"):
             source = (REPO_ROOT / dockerfile).read_text(encoding="utf-8")

tests/test_tts_config.py

@@ -10,6 +10,7 @@ from tts_config import (
     distribute_samples,
     engines_for_language,
     language_for_engine,
+    normalize_english_accent,
     normalize_tts_mode,
     quality_for_engines,
 )
@@ -51,6 +52,12 @@ class TtsConfigTests(unittest.TestCase):
     def test_invalid_mode_falls_back_to_four_provider_route(self) -> None:
         self.assertEqual(normalize_tts_mode("unknown"), "hybrid")
 
+    def test_english_accent_aliases_and_non_english_fallback(self) -> None:
+        self.assertEqual(normalize_english_accent("Australia", "en"), "australian")
+        self.assertEqual(normalize_english_accent("new-zealand", "en_US"), "new_zealand")
+        self.assertEqual(normalize_english_accent("unknown", "en"), "mixed")
+        self.assertEqual(normalize_english_accent("australian", "fr"), "mixed")
+
     def test_common_language_aliases_use_model_catalog_ids(self) -> None:
         self.assertEqual(language_for_engine(ENGINE_OMNIVOICE, "ar"), "arb")
         self.assertEqual(language_for_engine(ENGINE_OMNIVOICE, "ne"), "npi")

tests/test_vue_ui.py

@@ -80,6 +80,14 @@ class VueTrainerUiTests(unittest.TestCase):
         self.assertIn('@scroll.passive="onConsoleScroll"', app)
         self.assertIn("Jump to latest", app)
 
+    def test_saved_positive_and_negative_cards_keep_stt_results_visible(self) -> None:
+        app = (REPO_ROOT / "frontend" / "src" / "TrainerApp.vue").read_text(encoding="utf-8")
+        types = (REPO_ROOT / "frontend" / "src" / "types.ts").read_text(encoding="utf-8")
+
+        self.assertEqual(app.count('v-if="item.transcript" class="transcript"'), 2)
+        self.assertEqual(app.count('v-if="item.auto_review_guided_transcript"'), 2)
+        self.assertIn("auto_review_guided_transcript?: string", types)
+
     def test_wake_word_card_uses_explicit_json_catalog_url(self) -> None:
         app = (REPO_ROOT / "frontend" / "src" / "TrainerApp.vue").read_text(encoding="utf-8")
         types = (REPO_ROOT / "frontend" / "src" / "types.ts").read_text(encoding="utf-8")
@@ -89,6 +97,15 @@ class VueTrainerUiTests(unittest.TestCase):
         self.assertNotIn("copyWakeWord(word.url)", app)
         self.assertIn("json_url?: string", types)
 
+    def test_wake_words_tab_exposes_esphome_manifest_urls(self) -> None:
+        app = (REPO_ROOT / "frontend" / "src" / "TrainerApp.vue").read_text(encoding="utf-8")
+        types = (REPO_ROOT / "frontend" / "src" / "types.ts").read_text(encoding="utf-8")
+
+        self.assertIn("ESPHome JSON", app)
+        self.assertIn("wordEsphomeJsonUrl", app)
+        self.assertIn("Copy ESPHome URL", app)
+        self.assertIn("esphome_json_url?: string", types)
+
     def test_runtime_packaging_uses_bundle_without_node(self) -> None:
         dockerfiles = [REPO_ROOT / "dockerfile", REPO_ROOT / "dockerfile.blackwell"]
         for dockerfile in dockerfiles:

train_wake_word

@@ -5,7 +5,7 @@ PROGPATH=$(realpath "$0")
 PROGDIR=$(dirname "${PROGPATH}")
 CLIDIR="${PROGDIR}/cli"
 
-KNOWN_ARGS=( samples batch-size training-steps data-dir cleanup-work-dir language tts-mode tts-voice-count )
+KNOWN_ARGS=( samples batch-size training-steps data-dir cleanup-work-dir language english-accent tts-mode tts-voice-count )
 source "${CLIDIR}/shell.functions"
 WAKE_WORD=${POSITIONAL_ARGS[0]}
 
@@ -19,6 +19,7 @@ if [ "${HELP}" == "true" ] || [ -z "${WAKE_WORD}" ] ; then
 Usage: train_wake_word [ --samples=<samples> ] [ --batch-size=<batch_size> ]
                        [ --training-steps=<steps> ] [ --cleanup-work-dir ]
                        [ --language=<lang> ]
+                       [ --english-accent=<accent> ]
                        [ --tts-mode=<modern|hybrid|piper> ]
                        [ --tts-voice-count=<voices> ]
                        <wake_word> [ <wake_word_title> ]
@@ -41,6 +42,8 @@ Options:
 --language:           Language for TTS voice selection (e.g. "en", "nl").
                       Default: ${DEFAULT_LANGUAGE}
 
+--english-accent:     English accent emphasis. Default: ${DEFAULT_ENGLISH_ACCENT}
+
 --tts-mode:           TTS source: modern (OmniVoice plus Qwen3/MOSS where
                       supported), hybrid (modern plus Piper), or piper.
                       Default: ${DEFAULT_TTS_MODE}
@@ -124,6 +127,7 @@ export GRPC_VERBOSITY=ERROR
     --samples=${SAMPLES} \
     --batch-size=${BATCH_SIZE} \
     --language="${LANGUAGE}" \
+    --english-accent="${ENGLISH_ACCENT}" \
     --tts-mode="${TTS_MODE}" \
     --tts-voice-count="${TTS_VOICE_COUNT}" \
     --data-dir="${DATA_DIR}" "${WAKE_WORD}"

trainer_server.py

@@ -39,6 +39,7 @@ ROOT_DIR = Path(__file__).resolve().parent
 
 from tts_config import (
     COMMON_OMNIVOICE_LANGUAGES,
+    DEFAULT_ENGLISH_ACCENT,
     DEFAULT_TTS_MODE,
     ENGINE_MOSS,
     ENGINE_OMNIVOICE,
@@ -47,6 +48,8 @@ from tts_config import (
     MOSS_LANGUAGES,
     OMNIVOICE_LANGUAGE_ALIASES,
     QWEN_LANGUAGES,
+    english_accent_options,
+    normalize_english_accent,
     normalize_tts_mode,
     parse_omnivoice_catalog,
     quality_for_engines,
@@ -115,6 +118,10 @@ TRAIN_CMD = os.environ.get(
 )
 DEFAULT_LANGUAGE = os.environ.get("MWW_LANGUAGE", "en")
 DEFAULT_SERVER_TTS_MODE = normalize_tts_mode(os.environ.get("MWW_TTS_MODE", DEFAULT_TTS_MODE))
+DEFAULT_SERVER_ENGLISH_ACCENT = normalize_english_accent(
+    os.environ.get("MWW_ENGLISH_ACCENT", DEFAULT_ENGLISH_ACCENT),
+    DEFAULT_LANGUAGE,
+)
 
 TAKES_PER_SPEAKER_DEFAULT = int(os.environ.get("REC_TAKES_PER_SPEAKER", "10"))
 SPEAKERS_TOTAL_DEFAULT = int(os.environ.get("REC_SPEAKERS_TOTAL", "1"))
@@ -157,6 +164,7 @@ AUTO_TRAIN_DEFAULT_CONFIG: Dict[str, Any] = {
     "enabled": False,
     "wake_phrase": "",
     "language": DEFAULT_LANGUAGE,
+    "english_accent": DEFAULT_SERVER_ENGLISH_ACCENT,
     "stt_engine": DEFAULT_STT_ENGINE,
     "minimum_transcript_chars": 2,
     "delete_confirmed_wakes": False,
@@ -213,6 +221,7 @@ STATE: Dict[str, Any] = {
     "raw_phrase": None,
     "safe_word": None,
     "language": DEFAULT_LANGUAGE,
+    "english_accent": DEFAULT_SERVER_ENGLISH_ACCENT,
     "tts_mode": DEFAULT_SERVER_TTS_MODE,
 
     # multi-speaker
@@ -581,6 +590,24 @@ def _metadata_int(value: Any) -> int | None:
         return None
 
 
+ESPHOME_MANIFEST_SUFFIX = ".esphome.json"
+ESPHOME_MANIFEST_KEYS = (
+    "type",
+    "wake_word",
+    "author",
+    "website",
+    "model",
+    "trained_languages",
+    "version",
+    "micro",
+)
+
+
+def _esphome_manifest(metadata: Dict[str, Any]) -> Dict[str, Any]:
+    """Return only fields accepted by ESPHome's micro_wake_word v2 schema."""
+    return {key: metadata[key] for key in ESPHOME_MANIFEST_KEYS if key in metadata}
+
+
 def _list_trained_wake_words(base_url: str = "") -> List[Dict[str, Any]]:
     _sync_trained_wake_word_artifacts()
     base = str(base_url or "").rstrip("/")
@@ -619,9 +646,11 @@ def _list_trained_wake_words(base_url: str = "") -> List[Dict[str, Any]]:
         recall = _metadata_float(calibration.get("recall"))
         false_accepts_per_hour = _metadata_float(calibration.get("false_accepts_per_hour"))
         json_url = f"/api/trained_wake_words/{quote(json_path.name)}"
+        esphome_json_url = f"/api/trained_wake_words/{quote(safe + ESPHOME_MANIFEST_SUFFIX)}"
         model_url = f"/api/trained_wake_words/{quote(model_path.name)}"
         if base:
             json_url = f"{base}{json_url}"
+            esphome_json_url = f"{base}{esphome_json_url}"
             model_url = f"{base}{model_url}"
 
         rows.append(
@@ -634,6 +663,7 @@ def _list_trained_wake_words(base_url: str = "") -> List[Dict[str, Any]]:
                 # New consumers should prefer the explicit `json_url` field.
                 "url": json_url,
                 "json_url": json_url,
+                "esphome_json_url": esphome_json_url,
                 "model_url": model_url,
                 "json_file": json_path.name,
                 "model_file": model_path.name,
@@ -766,6 +796,9 @@ def _normalize_auto_train_config(values: Dict[str, Any] | None, *, base: Dict[st
         "enabled": _config_bool(source.get("enabled")),
         "wake_phrase": str(source.get("wake_phrase") or "").strip(),
         "language": language,
+        "english_accent": normalize_english_accent(
+            source.get("english_accent"), language
+        ),
         "stt_engine": _normalize_stt_engine(source.get("stt_engine")),
         "minimum_transcript_chars": _bounded_int(source.get("minimum_transcript_chars"), 2, 1, 100),
         "delete_confirmed_wakes": _config_bool(source.get("delete_confirmed_wakes")),
@@ -1704,6 +1737,9 @@ def _start_auto_training() -> Dict[str, Any]:
     safe_word = safe_name(wake_phrase)
     available_languages = _available_languages()
     language = _normalize_language(str(config.get("language") or DEFAULT_LANGUAGE))
+    english_accent = normalize_english_accent(
+        config.get("english_accent"), language
+    )
     tts_mode = _resolve_tts_mode_for_language(
         DEFAULT_SERVER_TTS_MODE,
         language,
@@ -1716,6 +1752,7 @@ def _start_auto_training() -> Dict[str, Any]:
             STATE["raw_phrase"] = wake_phrase
             STATE["safe_word"] = safe_word
             STATE["language"] = language
+            STATE["english_accent"] = english_accent
             STATE["tts_mode"] = tts_mode
             STATE["training"]["running"] = True
     with AUTO_TRAIN_LOCK:
@@ -1723,7 +1760,9 @@ def _start_auto_training() -> Dict[str, Any]:
         AUTO_TRAIN_RUNTIME["training_pending_consumed"] = int(AUTO_TRAIN_STATE.get("pending_negative_count") or 0)
         _save_auto_train_state_locked()
     try:
-        _start_training_thread(safe_word, language, True, True, tts_mode)
+        _start_training_thread(
+            safe_word, language, True, True, tts_mode, english_accent
+        )
     except Exception as exc:
         with STATE_LOCK:
             STATE["training"]["running"] = False
@@ -1733,6 +1772,7 @@ def _start_auto_training() -> Dict[str, Any]:
         "started": True,
         "safe_word": safe_word,
         "language": language,
+        "english_accent": english_accent,
         "tts_mode": tts_mode,
     }
 
@@ -2690,6 +2730,9 @@ def _sample_item_from_path(audio_path: Path, bucket: str) -> Dict[str, Any]:
         "message": meta.get("message") or "",
         "transcript": meta.get("transcript") or "",
         "transcribed_at": meta.get("transcribed_at") or "",
+        "auto_review_guided_transcript": meta.get("auto_review_guided_transcript") or "",
+        "auto_review_stt_engine": meta.get("auto_review_stt_engine") or "",
+        "auto_review_stt_model": meta.get("auto_review_stt_model") or "",
         "auto_negative": bool(meta.get("auto_negative")),
         "auto_positive": bool(meta.get("auto_positive")),
         "auto_review_reason": meta.get("auto_review_reason") or "",
@@ -3049,11 +3092,19 @@ def _start_training_thread(
     allow_no_personal: bool,
     auto_run: bool,
     tts_mode: str,
+    english_accent: str = DEFAULT_SERVER_ENGLISH_ACCENT,
 ) -> threading.Thread:
     global TRAINING_THREAD
     thread = threading.Thread(
         target=_run_training_background,
-        args=(safe_word, language, allow_no_personal, auto_run, tts_mode),
+        args=(
+            safe_word,
+            language,
+            allow_no_personal,
+            auto_run,
+            tts_mode,
+            english_accent,
+        ),
         daemon=True,
         name="wake-word-training",
     )
@@ -3097,10 +3148,12 @@ def _run_training_background(
     allow_no_personal: bool,
     auto_run: bool = False,
     tts_mode: str = DEFAULT_SERVER_TTS_MODE,
+    english_accent: str = DEFAULT_SERVER_ENGLISH_ACCENT,
 ):
     global TRAINING_PROCESS, TRAINING_THREAD
     language = (language or DEFAULT_LANGUAGE).strip().lower() or DEFAULT_LANGUAGE
     tts_mode = normalize_tts_mode(tts_mode)
+    english_accent = normalize_english_accent(english_accent, language)
     rc = 999
     proc: subprocess.Popen | None = None
     with STATE_LOCK:
@@ -3146,7 +3199,12 @@ def _run_training_background(
             except Exception as error:
                 _append_train_log(f"⚠️ Piper is unavailable for hybrid mode; using modern TTS only: {error}")
 
-        command_args = [f"--language={language}", f"--tts-mode={tts_mode}", safe_word]
+        command_args = [
+            f"--language={language}",
+            f"--english-accent={english_accent}",
+            f"--tts-mode={tts_mode}",
+            safe_word,
+        ]
         if wake_word_title:
             command_args.append(wake_word_title)
         cmd_str = f"{TRAIN_CMD} " + " ".join(shlex.quote(argument) for argument in command_args)
@@ -3156,6 +3214,8 @@ def _run_training_background(
 
         _append_train_log("===== Training (train_wake_word) =====")
         _append_train_log(f"→ Running: {cmd_str}")
+        if language == "en":
+            _append_train_log(f"→ English accent emphasis: {english_accent}")
 
         with open(log_path, "a", encoding="utf-8") as lf:
             proc = subprocess.Popen(
@@ -3387,6 +3447,9 @@ def start_session(payload: Dict[str, Any]):
         language,
         available_languages,
     )
+    english_accent = normalize_english_accent(
+        payload.get("english_accent", DEFAULT_SERVER_ENGLISH_ACCENT), language
+    )
 
     speakers_total = max(1, min(10, speakers_total))
     takes_per_speaker = max(1, min(50, takes_per_speaker))
@@ -3395,6 +3458,7 @@ def start_session(payload: Dict[str, Any]):
         STATE["raw_phrase"] = raw
         STATE["safe_word"] = safe
         STATE["language"] = language
+        STATE["english_accent"] = english_accent
         STATE["tts_mode"] = tts_mode
         STATE["speakers_total"] = speakers_total
         STATE["takes_per_speaker"] = takes_per_speaker
@@ -3408,6 +3472,7 @@ def start_session(payload: Dict[str, Any]):
         "raw_phrase": raw,
         "safe_word": safe,
         "language": language,
+        "english_accent": english_accent,
         "tts_mode": tts_mode,
         "speakers_total": speakers_total,
         "takes_per_speaker": takes_per_speaker,
@@ -3415,6 +3480,7 @@ def start_session(payload: Dict[str, Any]):
         "takes_received": len(takes),
         "takes": takes,
         "available_languages": available_languages,
+        "available_english_accents": english_accent_options(),
         "personal_dir": str(PERSONAL_DIR),
         "data_dir": str(DATA_DIR),
     }
@@ -3444,6 +3510,9 @@ def stop_session():
         STATE["training"]["safe_word"] = None
         training = dict(STATE["training"])
         language = _normalize_language(STATE["language"])
+        english_accent = normalize_english_accent(
+            STATE.get("english_accent"), language
+        )
         tts_mode = normalize_tts_mode(STATE.get("tts_mode"))
     return {
         "ok": True,
@@ -3452,11 +3521,13 @@ def stop_session():
         "raw_phrase": None,
         "safe_word": None,
         "language": language,
+        "english_accent": english_accent,
         "tts_mode": tts_mode,
         "takes_received": len(takes),
         "takes": list(takes),
         "training": training,
         "available_languages": available_languages,
+        "available_english_accents": english_accent_options(),
     }
 
 
@@ -3467,13 +3538,18 @@ def get_session():
     with STATE_LOCK:
         current_language = _normalize_language(STATE["language"])
         current_tts_mode = normalize_tts_mode(STATE.get("tts_mode"))
+        current_english_accent = normalize_english_accent(
+            STATE.get("english_accent"), current_language
+        )
         STATE["language"] = current_language
+        STATE["english_accent"] = current_english_accent
         STATE["tts_mode"] = current_tts_mode
         return {
             "ok": True,
             "raw_phrase": STATE["raw_phrase"],
             "safe_word": STATE["safe_word"],
             "language": current_language,
+            "english_accent": current_english_accent,
             "tts_mode": current_tts_mode,
             "speakers_total": STATE["speakers_total"],
             "takes_per_speaker": STATE["takes_per_speaker"],
@@ -3481,6 +3557,7 @@ def get_session():
             "takes": list(takes),
             "training": dict(STATE["training"]),
             "available_languages": available_languages,
+            "available_english_accents": english_accent_options(),
         }
 
 
@@ -3960,6 +4037,24 @@ def trained_wake_word_artifact(filename: str):
     if not safe_filename or Path(safe_filename).suffix.lower() not in {".json", ".tflite"}:
         return JSONResponse({"ok": False, "error": "Unsupported wake word artifact."}, status_code=400)
     _sync_trained_wake_word_artifacts()
+    if safe_filename.endswith(ESPHOME_MANIFEST_SUFFIX):
+        source_stem = safe_filename[: -len(ESPHOME_MANIFEST_SUFFIX)]
+        source_path = TRAINED_WAKE_WORDS_DIR / f"{source_stem}.json"
+        if not source_stem or not source_path.is_file():
+            return JSONResponse({"ok": False, "error": "Wake word artifact not found."}, status_code=404)
+        try:
+            metadata = json.loads(source_path.read_text(encoding="utf-8"))
+        except Exception:
+            return JSONResponse({"ok": False, "error": "Wake word package is invalid."}, status_code=422)
+        if not isinstance(metadata, dict):
+            return JSONResponse({"ok": False, "error": "Wake word package is invalid."}, status_code=422)
+        model_name = Path(str(metadata.get("model") or f"{source_stem}.tflite")).name
+        if not (TRAINED_WAKE_WORDS_DIR / model_name).is_file():
+            return JSONResponse({"ok": False, "error": "Wake word model not found."}, status_code=404)
+        return JSONResponse(
+            _esphome_manifest(metadata),
+            headers={"Cache-Control": "no-store, max-age=0"},
+        )
     artifact_path = TRAINED_WAKE_WORDS_DIR / safe_filename
     if not artifact_path.exists() or not artifact_path.is_file():
         return JSONResponse({"ok": False, "error": "Wake word artifact not found."}, status_code=404)
@@ -3975,6 +4070,9 @@ def train_now(payload: Dict[str, Any] = None):
     with STATE_LOCK:
         safe_word = STATE["safe_word"]
         language = (STATE.get("language") or DEFAULT_LANGUAGE)
+        english_accent = normalize_english_accent(
+            STATE.get("english_accent"), language
+        )
         tts_mode = normalize_tts_mode(STATE.get("tts_mode"))
         takes_received = int(STATE["takes_received"])
         speakers_total = int(STATE["speakers_total"])
@@ -4004,7 +4102,14 @@ def train_now(payload: Dict[str, Any] = None):
     with STATE_LOCK:
         STATE["training"]["running"] = True
     try:
-        _start_training_thread(safe_word, language, allow_no_personal, False, tts_mode)
+        _start_training_thread(
+            safe_word,
+            language,
+            allow_no_personal,
+            False,
+            tts_mode,
+            english_accent,
+        )
     except Exception as exc:
         with STATE_LOCK:
             STATE["training"]["running"] = False
@@ -4018,6 +4123,7 @@ def train_now(payload: Dict[str, Any] = None):
         "started": True,
         "safe_word": safe_word,
         "language": language,
+        "english_accent": english_accent,
         "tts_mode": tts_mode,
         "personal_samples_used": takes_received > 0,
         "allow_no_personal": allow_no_personal,

tts_config.py

@@ -16,6 +16,26 @@ TTS_MODE_PIPER = "piper"
 TTS_MODES = (TTS_MODE_MODERN, TTS_MODE_HYBRID, TTS_MODE_PIPER)
 DEFAULT_TTS_MODE = TTS_MODE_HYBRID
 
+# English is one TTS language, while these values control the accent mix used
+# by providers that can follow a style instruction or clone a reference. The
+# remaining providers continue contributing their available English voices.
+DEFAULT_ENGLISH_ACCENT = "mixed"
+ENGLISH_ACCENTS = {
+    "mixed": "Mixed English",
+    "australian": "Australian",
+    "american": "American",
+    "british": "British",
+    "canadian": "Canadian",
+    "irish": "Irish",
+    "scottish": "Scottish",
+    "new_zealand": "New Zealand",
+    "indian": "Indian",
+    "south_african": "South African",
+}
+MIXED_ENGLISH_ACCENTS = tuple(
+    code for code in ENGLISH_ACCENTS if code != DEFAULT_ENGLISH_ACCENT
+)
+
 ENGINE_OMNIVOICE = "omnivoice"
 ENGINE_QWEN3 = "qwen3"
 ENGINE_MOSS = "moss"
@@ -157,6 +177,39 @@ def normalize_tts_mode(value: object) -> str:
     return token if token in TTS_MODES else DEFAULT_TTS_MODE
 
 
+def normalize_english_accent(value: object, language: object = "en") -> str:
+    """Return a supported English accent emphasis or the mixed default."""
+
+    language_code = str(language or "en").strip().lower().replace("-", "_")
+    if language_code.split("_", 1)[0] != "en":
+        return DEFAULT_ENGLISH_ACCENT
+
+    token = str(value or DEFAULT_ENGLISH_ACCENT).strip().lower().replace("-", "_").replace(" ", "_")
+    aliases = {
+        "all": "mixed",
+        "none": "mixed",
+        "us": "american",
+        "usa": "american",
+        "uk": "british",
+        "gb": "british",
+        "australia": "australian",
+        "canada": "canadian",
+        "ireland": "irish",
+        "scotland": "scottish",
+        "new_zealand_english": "new_zealand",
+        "south_africa": "south_african",
+    }
+    token = aliases.get(token, token)
+    return token if token in ENGLISH_ACCENTS else DEFAULT_ENGLISH_ACCENT
+
+
+def english_accent_options() -> list[dict[str, str]]:
+    return [
+        {"code": code, "label": label}
+        for code, label in ENGLISH_ACCENTS.items()
+    ]
+
+
 def language_for_engine(engine: str, language: str) -> str:
     code = str(language or "en").strip().lower().replace("-", "_")
     if engine == ENGINE_OMNIVOICE: