Release NVIDIA WakeWord Trainer v26

VERSION

@@ -1 +1 @@
-22
+26

WHATS_NEW.md

@@ -1,19 +1 @@
-- Fixed the training console so scrolling up pauses auto-follow and provides a Jump to latest control.
-- Fixed trained wake-word cards and Copy URL to use the explicit JSON package URL instead of producing `undefined`.
-- Replaced the 128-profile clone pipeline with direct final-corpus generation from Qwen, OmniVoice, and Piper; MOSS now uses a different accepted carrier for each take.
-- Added strict provider-specific rejection for static, broadband/high-frequency noise, silence, clipping, excessive duration/rambling, and exact duplicate audio.
-- Capped Qwen and MOSS decoding for a single short utterance, fixed OmniVoice to bounded wake-phrase durations, and let safer providers fill every rejected share.
-- Expanded Qwen to 18,750 balanced combinations across gender, age, pitch, delivery, timbre, pace, and vocal weight before an instruction repeats.
-- Shifted the reactive trainer UI from blue-black surfaces to Tater's graphite-grey and orange visual theme.
-- Rebuilt the trainer interface as a reactive Vue 3 + TypeScript application using the same typed UI pattern as Tater.
-- Preserved session setup, multilingual TTS routing, sample review/import/trim, Auto Training, secure Tater pairing, live logs, and wake-word publishing in the new component-driven UI.
-- Updated the standard CUDA and Blackwell Dockerfiles to copy the complete prebuilt UI bundle; Node.js is not installed or required in the runtime image.
-- Made OmniVoice, Qwen3-TTS, MOSS-TTS-Nano, and Piper the recommended four-provider route where a compatible Piper model is present.
-- Added the live 646-language OmniVoice catalog, language quality tiers, exact per-engine routing, and normalized acoustic QA.
-- Added persistent per-engine environments and Hugging Face caches that keep conflicting TTS dependencies separate from wake-word training.
-- Improved automatic review accuracy for short wake phrases that STT initially hears as similar-sounding words.
-- Added a conservative Faster Whisper confirmation pass that uses the currently configured wake phrase only when the unbiased transcript is already phonetically close.
-- Kept unconfirmed close transcripts in the manual review inbox instead of allowing them to become harmful negative training samples.
-- Added visible guided-transcript and review-reason details, plus retry support for ambiguous clips through Review Now.
-- Locked the wake phrase, language, and TTS route while a session is active, and added Stop Session with clean full-process-tree training cancellation.
-- Added a Data tab with per-dataset disk usage and file counts, plus confirmed, training-safe deletion for recordings, downloads, generated caches, speech models, and training results.
+- Restored MOSS-TTS-Nano generation with the current voice-cloning API so all four providers contribute to the final training corpus.

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
@@ -43,7 +44,7 @@ from tts_config import (  # noqa: E402
 )
 
 
-GENERATOR_VERSION = "modern-tts-v15-four-provider-direct-corpus-safe-limits"
+GENERATOR_VERSION = "modern-tts-v16-four-provider-direct-corpus-safe-limits"
 VOICE_BANK_VERSION = "modern-tts-voice-bank-v1-native-random-qualified-single-utterance"
 COMPATIBLE_VOICE_BANK_VERSIONS = {
     VOICE_BANK_VERSION,
@@ -65,6 +66,8 @@ 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 +125,38 @@ 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:
@@ -1247,7 +1282,9 @@ 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 +1295,7 @@ class Generator:
                 "-hide_banner",
                 "-loglevel",
                 "error",
+                "-nostdin",
                 "-y",
                 "-i",
                 str(path),
@@ -1272,11 +1310,22 @@ 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
+                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)
@@ -1284,6 +1333,16 @@ class Generator:
                     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:

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,

frontend/src/TrainerApp.vue

@@ -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();
@@ -265,6 +266,7 @@ 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 +303,11 @@ 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/types.ts

@@ -31,6 +31,11 @@ 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 {
@@ -76,6 +81,8 @@ export interface WakeWordItem extends JsonRecord {
   url?: string;
   json_url?: string;
   jsonUrl?: string;
+  esphome_json_url?: string;
+  esphomeJsonUrl?: string;
   model_url?: string;
   modelUrl?: string;
 }

static/ui/trainer-ui.js

@@ -3983,44 +3983,59 @@ var hc = {
 }, du = { class: "row" }, fu = {
 	key: 0,
 	class: "pill warning"
-}, pu = ["src"], mu = ["onClick"], hu = ["onClick"], gu = ["disabled", "onClick"], _u = {
+}, pu = {
+	key: 0,
+	class: "transcript"
+}, mu = {
+	key: 1,
+	class: "transcript"
+}, hu = ["src"], gu = ["onClick"], _u = ["onClick"], vu = ["disabled", "onClick"], yu = {
 	key: 2,
 	class: "pagination"
-}, vu = ["disabled"], yu = ["disabled"], bu = { class: "panel" }, xu = { class: "dropzone" }, Su = ["disabled"], Cu = { class: "progress-card" }, wu = { class: "progress-track" }, Tu = { class: "hero data-hero" }, Eu = { class: "pill hero-pill" }, Du = { class: "panel" }, Ou = { class: "panel-head" }, ku = ["disabled"], Au = { class: "stats" }, ju = { class: "format-value" }, Mu = {
+}, bu = ["disabled"], xu = ["disabled"], Su = { class: "panel" }, Cu = { class: "dropzone" }, wu = ["disabled"], Tu = { class: "progress-card" }, Eu = { class: "progress-track" }, Du = { class: "hero data-hero" }, Ou = { class: "pill hero-pill" }, ku = { class: "panel" }, Au = { class: "panel-head" }, ju = ["disabled"], Mu = { class: "stats" }, Nu = { class: "format-value" }, Pu = {
 	key: 0,
 	class: "data-warning"
-}, Nu = { class: "panel-head" }, Pu = { class: "number" }, Fu = { class: "data-list" }, Iu = { class: "data-copy" }, Lu = { class: "data-title" }, Ru = {
+}, Fu = { class: "panel-head" }, Iu = { class: "number" }, Lu = { class: "data-list" }, Ru = { class: "data-copy" }, zu = { class: "data-title" }, Bu = {
 	key: 0,
 	class: "data-note"
-}, zu = { class: "data-usage" }, Bu = ["disabled", "onClick"], Vu = {
+}, Vu = { class: "data-usage" }, Hu = ["disabled", "onClick"], Uu = {
 	key: 0,
 	class: "panel empty-state"
-}, Hu = { class: "hero firmware-hero" }, Uu = { class: "panel" }, Wu = { class: "panel-head" }, Gu = ["disabled"], Ku = {
+}, Wu = { class: "hero firmware-hero" }, Gu = { class: "panel" }, Ku = { class: "panel-head" }, qu = ["disabled"], Ju = {
 	key: 0,
 	class: "empty-state"
-}, qu = {
+}, Yu = {
 	key: 1,
 	class: "word-list"
-}, Ju = ["href"], Yu = {
+}, Xu = ["href"], Zu = {
 	key: 1,
 	class: "muted"
-}, Xu = ["href"], Zu = { class: "meta-row" }, Qu = { key: 0 }, $u = { key: 1 }, ed = { key: 2 }, td = ["disabled", "onClick"], nd = {
+}, Qu = ["href"], $u = { class: "meta-row" }, ed = { key: 0 }, td = { key: 1 }, nd = { key: 2 }, rd = ["disabled", "onClick"], id = { class: "panel compatibility-panel" }, ad = {
+	key: 0,
+	class: "empty-state"
+}, od = {
+	key: 1,
+	class: "word-list"
+}, sd = ["href"], cd = {
+	key: 1,
+	class: "muted"
+}, ld = ["disabled", "onClick"], ud = {
 	class: "modal console-modal",
 	role: "dialog",
 	"aria-modal": "true",
 	"aria-label": "Training console"
-}, rd = { class: "modal-head" }, id = { class: "row console-actions" }, ad = {
+}, dd = { class: "modal-head" }, fd = { class: "row console-actions" }, pd = {
 	class: "modal link-modal",
 	role: "dialog",
 	"aria-modal": "true",
 	"aria-label": "Link Tater"
-}, od = { class: "modal-head" }, sd = {
+}, md = { class: "modal-head" }, hd = {
 	key: 0,
 	class: "link-success"
-}, cd = {
+}, gd = {
 	key: 1,
 	class: "stack"
-}, ld = { class: "field" }, ud = { class: "field" }, dd = ["disabled"], fd = "/static/images/tater-wake-word-trainer.png", pd = 50, md = /* @__PURE__ */ fr({
+}, _d = { class: "field" }, vd = { class: "field" }, yd = ["disabled"], bd = "/static/images/tater-wake-word-trainer.png", xd = 50, Sd = /* @__PURE__ */ fr({
 	__name: "TrainerApp",
 	setup(e) {
 		let t = /* @__PURE__ */ F(null), n = /* @__PURE__ */ F(null), r = /* @__PURE__ */ F(!0), i = /* @__PURE__ */ F(""), a = /* @__PURE__ */ F(""), o = /* @__PURE__ */ F(!1), s = [
@@ -4056,8 +4071,8 @@ var hc = {
 			}
 		], c = Y(() => {
 			let e = X.samplePage[X.sampleBucket];
-			return As.value.slice(e * pd, (e + 1) * pd);
-		}), l = Y(() => Math.max(1, Math.ceil(As.value.length / pd))), u = Y(() => X.auto.state || {}), d = Y(() => X.auto.runtime || {}), f = Y(() => {
+			return As.value.slice(e * xd, (e + 1) * xd);
+		}), l = Y(() => Math.max(1, Math.ceil(As.value.length / xd))), u = Y(() => X.auto.state || {}), d = Y(() => X.auto.runtime || {}), f = Y(() => {
 			let e = u.value, t = [];
 			return e.last_review_result && t.push(`Last review: ${String(e.last_review_result).replaceAll("_", " ")}`), e.last_review_file && t.push(String(e.last_review_file)), e.last_review_transcript && t.push(`STT: “${e.last_review_transcript}”`), e.last_review_error && t.push(`Error: ${e.last_review_error}`), e.last_stt_engine && t.push(`STT engine: ${String(e.last_stt_engine).replaceAll("_", " ")}`), e.last_notify_at && t.push(e.last_notify_error ? `Publish failed: ${e.last_notify_error}` : `Wake word published ${uc(e.last_notify_at)}`), t.join(" · ") || "No automatic review has run yet.";
 		}), p = Y(() => X.training.running ? {
@@ -4156,18 +4171,21 @@ var hc = {
 			return String(e.json_url || e.url || e.jsonUrl || "");
 		}
 		function re(e) {
-			return String(e.model_url || e.modelUrl || "");
+			return String(e.esphome_json_url || e.esphomeJsonUrl || "");
 		}
 		function E(e) {
+			return String(e.model_url || e.modelUrl || "");
+		}
+		function ie(e) {
 			let t = e.trim().toLowerCase();
 			return /^(✓|✅)|success|finished/.test(t) ? "success" : /^(✗|❌)|error|failed|traceback/.test(t) ? "error" : /^(⚠|warning)/.test(t) ? "warning" : /^={4,}|^-----|^=====/.test(t) ? "heading" : "";
 		}
 		return (e, d) => (U(), W("div", Dc, [
-			d[113] ||= G("div", {
+			d[118] ||= G("div", {
 				class: "ambient ambient-one",
 				"aria-hidden": "true"
 			}, null, -1),
-			d[114] ||= G("div", {
+			d[119] ||= G("div", {
 				class: "ambient ambient-two",
 				"aria-hidden": "true"
 			}, null, -1),
@@ -4175,7 +4193,7 @@ var hc = {
 				class: "brand-mark",
 				"aria-hidden": "true"
 			}, [G("img", {
-				src: fd,
+				src: bd,
 				alt: ""
 			})]), d[44] ||= G("div", null, [
 				G("span", { class: "eyebrow" }, "Tater tools"),
@@ -4518,46 +4536,48 @@ var hc = {
 						class: "audio-card"
 					}, [
 						G("header", null, [G("div", null, [G("strong", null, k(e.saved_as), 1), G("small", null, k(ne(e)), 1)]), G("div", du, [e.trimmed ? (U(), W("span", fu, "Trimmed")) : q("", !0), G("span", { class: O(["pill", I(X).sampleBucket === "personal" ? "success" : "error"]) }, k(I(X).sampleBucket === "personal" ? "Positive" : "Negative"), 3)])]),
+						e.transcript ? (U(), W("div", pu, [d[94] ||= G("b", null, "STT", -1), da(" " + k(e.transcript), 1)])) : q("", !0),
+						e.auto_review_guided_transcript ? (U(), W("div", mu, [d[95] ||= G("b", null, "Guided wake check", -1), da(" " + k(e.auto_review_guided_transcript), 1)])) : q("", !0),
 						G("audio", {
 							controls: "",
 							preload: "none",
 							src: I(mc)(e, I(X).sampleBucket)
-						}, null, 8, pu),
+						}, null, 8, hu),
 						G("footer", null, [G("span", null, k(I(fc)(e.final_format)), 1), G("div", null, [
 							G("button", {
 								type: "button",
 								onClick: (t) => S(e, I(X).sampleBucket)
-							}, "Trim", 8, mu),
+							}, "Trim", 8, gu),
 							e.trimmed ? (U(), W("button", {
 								key: 0,
 								type: "button",
 								onClick: (t) => I(Js)(e, I(X).sampleBucket)
-							}, "Revert", 8, hu)) : q("", !0),
+							}, "Revert", 8, _u)) : q("", !0),
 							G("button", {
 								type: "button",
 								class: "button danger ghost",
 								disabled: I(Z)("review"),
 								onClick: (t) => I(qs)(e, I(X).sampleBucket)
-							}, "Remove", 8, gu)
+							}, "Remove", 8, vu)
 						])])
 					]))), 128))])) : (U(), W("div", lu, "No " + k(I(X).sampleBucket) + " samples saved yet.", 1)),
-					l.value > 1 ? (U(), W("div", _u, [
+					l.value > 1 ? (U(), W("div", yu, [
 						G("button", {
 							type: "button",
 							disabled: I(X).samplePage[I(X).sampleBucket] === 0,
 							onClick: d[32] ||= (e) => I(X).samplePage[I(X).sampleBucket]--
-						}, "Previous", 8, vu),
+						}, "Previous", 8, bu),
 						G("span", null, "Page " + k(I(X).samplePage[I(X).sampleBucket] + 1) + " of " + k(l.value), 1),
 						G("button", {
 							type: "button",
 							disabled: I(X).samplePage[I(X).sampleBucket] >= l.value - 1,
 							onClick: d[33] ||= (e) => I(X).samplePage[I(X).sampleBucket]++
-						}, "Next", 8, yu)
+						}, "Next", 8, xu)
 					])) : q("", !0)
 				]),
-				G("section", bu, [
-					d[95] ||= G("header", { class: "panel-head" }, [G("div", { class: "number" }, "2"), G("div", null, [G("h3", null, "Manual sample import"), G("p", null, "Optional seed recordings are normalized to the trainer’s required WAV format.")])], -1),
-					G("label", xu, [
+				G("section", Su, [
+					d[97] ||= G("header", { class: "panel-head" }, [G("div", { class: "number" }, "2"), G("div", null, [G("h3", null, "Manual sample import"), G("p", null, "Optional seed recordings are normalized to the trainer’s required WAV format.")])], -1),
+					G("label", Cu, [
 						G("input", {
 							ref_key: "uploadInput",
 							ref: t,
@@ -4566,7 +4586,7 @@ var hc = {
 							accept: "audio/*,.wav,.mp3,.m4a,.flac,.ogg,.aac,.webm,.opus",
 							onChange: d[34] ||= (...e) => I(Us) && I(Us)(...e)
 						}, null, 544),
-						d[94] ||= G("span", null, [G("strong", null, "Choose one or many audio files"), G("small", null, "WAV, MP3, M4A, FLAC, OGG, AAC, OPUS, and WEBM")], -1),
+						d[96] ||= G("span", null, [G("strong", null, "Choose one or many audio files"), G("small", null, "WAV, MP3, M4A, FLAC, OGG, AAC, OPUS, and WEBM")], -1),
 						G("b", null, k(I(X).selectedFiles.length ? `${I(X).selectedFiles.length} selected` : "Browse"), 1)
 					]),
 					G("button", {
@@ -4574,106 +4594,121 @@ var hc = {
 						class: "button primary",
 						disabled: !I(X).session.safe_word || !I(X).selectedFiles.length || I(Z)("upload"),
 						onClick: d[35] ||= (e) => I(Gs)(t.value)
-					}, k(I(Z)("upload") ? "Uploading…" : "Upload selected samples"), 9, Su),
-					G("div", Cu, [
+					}, k(I(Z)("upload") ? "Uploading…" : "Upload selected samples"), 9, wu),
+					G("div", Tu, [
 						G("div", null, [G("strong", null, k(I(X).uploadLabel), 1), G("span", null, k(I(X).uploadProgress) + "%", 1)]),
-						G("div", wu, [G("i", { style: fe({ width: `${I(X).uploadProgress}%` }) }, null, 4)]),
+						G("div", Eu, [G("i", { style: fe({ width: `${I(X).uploadProgress}%` }) }, null, 4)]),
 						G("small", null, k(I(X).uploadDetail), 1)
 					])
 				])
 			], 64)) : I(X).activeView === "data" ? (U(), W(V, { key: 4 }, [
-				G("section", Tu, [d[96] ||= G("div", null, [
+				G("section", Du, [d[98] ||= G("div", null, [
 					G("span", { class: "eyebrow" }, "Local storage"),
 					G("h2", null, "Data Management"),
 					G("p", null, "See exactly what the trainer has downloaded, generated, recorded, and produced.")
-				], -1), G("span", Eu, k(I(dc)(I(X).managedData.total_size_bytes)) + " total", 1)]),
-				G("section", Du, [
-					G("header", Ou, [
-						d[97] ||= G("div", { class: "number" }, "i", -1),
-						d[98] ||= G("div", null, [G("h3", null, "Trainer storage"), G("p", null, "Deleting an item is permanent. Required downloads and generated caches will be rebuilt the next time training needs them.")], -1),
+				], -1), G("span", Ou, k(I(dc)(I(X).managedData.total_size_bytes)) + " total", 1)]),
+				G("section", ku, [
+					G("header", Au, [
+						d[99] ||= G("div", { class: "number" }, "i", -1),
+						d[100] ||= G("div", null, [G("h3", null, "Trainer storage"), G("p", null, "Deleting an item is permanent. Required downloads and generated caches will be rebuilt the next time training needs them.")], -1),
 						G("button", {
 							type: "button",
 							disabled: I(Z)("data") || I(Z)("data-delete"),
 							onClick: d[36] ||= (e) => I(rc)()
-						}, k(I(Z)("data") ? "Scanning…" : "Refresh sizes"), 9, ku)
+						}, k(I(Z)("data") ? "Scanning…" : "Refresh sizes"), 9, ju)
 					]),
-					G("div", Au, [
-						G("article", null, [d[99] ||= G("span", null, "Space used", -1), G("strong", ju, k(I(dc)(I(X).managedData.total_size_bytes)), 1)]),
-						G("article", null, [d[100] ||= G("span", null, "Files", -1), G("strong", null, k(Number(I(X).managedData.total_file_count || 0).toLocaleString()), 1)]),
-						G("article", null, [d[101] ||= G("span", null, "Individual items", -1), G("strong", null, k(I(X).managedData.items.length), 1)])
+					G("div", Mu, [
+						G("article", null, [d[101] ||= G("span", null, "Space used", -1), G("strong", Nu, k(I(dc)(I(X).managedData.total_size_bytes)), 1)]),
+						G("article", null, [d[102] ||= G("span", null, "Files", -1), G("strong", null, k(Number(I(X).managedData.total_file_count || 0).toLocaleString()), 1)]),
+						G("article", null, [d[103] ||= G("span", null, "Individual items", -1), G("strong", null, k(I(X).managedData.items.length), 1)])
 					]),
-					I(X).training.running ? (U(), W("p", Mu, "Stop the active training session before deleting data.")) : q("", !0)
+					I(X).training.running ? (U(), W("p", Pu, "Stop the active training session before deleting data.")) : q("", !0)
 				]),
 				(U(!0), W(V, null, Lr(g.value, (e, t) => (U(), W("section", {
 					key: e.name,
 					class: "panel data-panel"
-				}, [G("header", Nu, [G("div", Pu, k(t + 1), 1), G("div", null, [G("h3", null, k(e.name), 1), G("p", null, k(e.items.length) + " separately managed item" + k(e.items.length === 1 ? "" : "s"), 1)])]), G("div", Fu, [(U(!0), W(V, null, Lr(e.items, (e) => (U(), W("article", {
+				}, [G("header", Fu, [G("div", Iu, k(t + 1), 1), G("div", null, [G("h3", null, k(e.name), 1), G("p", null, k(e.items.length) + " separately managed item" + k(e.items.length === 1 ? "" : "s"), 1)])]), G("div", Lu, [(U(!0), W(V, null, Lr(e.items, (e) => (U(), W("article", {
 					key: e.id,
 					class: O(["data-row", { empty: !e.file_count }])
 				}, [
-					G("div", Iu, [
-						G("div", Lu, [G("strong", null, k(e.label), 1), G("code", null, k(e.location), 1)]),
+					G("div", Ru, [
+						G("div", zu, [G("strong", null, k(e.label), 1), G("code", null, k(e.location), 1)]),
 						G("small", null, k(e.description), 1),
-						e.rebuild_note ? (U(), W("span", Ru, k(e.rebuild_note), 1)) : q("", !0)
+						e.rebuild_note ? (U(), W("span", Bu, k(e.rebuild_note), 1)) : q("", !0)
 					]),
-					G("div", zu, [G("strong", null, k(I(dc)(e.size_bytes)), 1), G("span", null, k(Number(e.file_count || 0).toLocaleString()) + " file" + k(e.file_count === 1 ? "" : "s"), 1)]),
+					G("div", Vu, [G("strong", null, k(I(dc)(e.size_bytes)), 1), G("span", null, k(Number(e.file_count || 0).toLocaleString()) + " file" + k(e.file_count === 1 ? "" : "s"), 1)]),
 					G("button", {
 						type: "button",
 						class: "button danger ghost",
 						disabled: !e.file_count || I(X).training.running || I(Z)("data") || I(Z)("data-delete"),
 						onClick: (t) => I(ic)(e)
-					}, k(I(Z)("data-delete") ? "Please wait…" : "Delete"), 9, Bu)
+					}, k(I(Z)("data-delete") ? "Please wait…" : "Delete"), 9, Hu)
 				], 2))), 128))])]))), 128)),
-				!I(Z)("data") && !I(X).managedData.items.length ? (U(), W("section", Vu, "No managed trainer data was found.")) : q("", !0)
+				!I(Z)("data") && !I(X).managedData.items.length ? (U(), W("section", Uu, "No managed trainer data was found.")) : q("", !0)
 			], 64)) : I(X).activeView === "firmware" ? (U(), W(V, { key: 5 }, [
-				G("section", Hu, [d[102] ||= G("div", null, [
+				G("section", Wu, [d[104] ||= G("div", null, [
 					G("span", { class: "eyebrow" }, "Wake-word catalog"),
 					G("h2", null, "Trained Wake Words"),
 					G("p", null, "Copy a local JSON package URL into Tater to switch every native satellite live.")
 				], -1), G("span", { class: O(["pill hero-pill", I(X).wakeWords.length ? "success" : "warning"]) }, k(I(X).wakeWords.length ? `${I(X).wakeWords.length} trained` : "Catalog empty"), 3)]),
-				d[105] ||= G("div", { class: "native-notice" }, [G("strong", null, "Tater Native"), G("span", null, "These packages include model metadata and a direct model URL for live satellite updates.")], -1),
-				G("section", Uu, [G("header", Wu, [
-					d[103] ||= G("div", { class: "number" }, "v1", -1),
-					d[104] ||= G("div", null, [G("h3", null, "Published model URLs"), G("p", null, "URLs stay local and are refreshed after each successful run.")], -1),
+				d[109] ||= G("div", { class: "native-notice" }, [G("strong", null, "Tater Native"), G("span", null, "These packages include model metadata and a direct model URL for live satellite updates.")], -1),
+				G("section", Gu, [G("header", Ku, [
+					d[105] ||= G("div", { class: "number" }, "v1", -1),
+					d[106] ||= G("div", null, [G("h3", null, "Published model URLs"), G("p", null, "URLs stay local and are refreshed after each successful run.")], -1),
 					G("button", {
 						type: "button",
 						disabled: I(Z)("firmware"),
 						onClick: d[37] ||= (e) => I(nc)()
-					}, "Refresh", 8, Gu)
-				]), I(X).wakeWords.length ? (U(), W("div", qu, [(U(!0), W(V, null, Lr(I(X).wakeWords, (e) => (U(), W("article", { key: e.key || T(e) }, [G("div", null, [
+					}, "Refresh", 8, qu)
+				]), I(X).wakeWords.length ? (U(), W("div", Yu, [(U(!0), W(V, null, Lr(I(X).wakeWords, (e) => (U(), W("article", { key: e.key || T(e) }, [G("div", null, [
 					G("strong", null, k(e.label || e.name || "Trained wake word"), 1),
 					T(e) ? (U(), W("a", {
 						key: 0,
 						href: T(e),
 						target: "_blank",
 						rel: "noreferrer"
-					}, "JSON · " + k(T(e)), 9, Ju)) : (U(), W("span", Yu, "JSON package URL unavailable")),
-					re(e) ? (U(), W("a", {
+					}, "JSON · " + k(T(e)), 9, Xu)) : (U(), W("span", Zu, "JSON package URL unavailable")),
+					E(e) ? (U(), W("a", {
 						key: 2,
-						href: re(e),
+						href: E(e),
 						target: "_blank",
 						rel: "noreferrer"
-					}, "Model · " + k(re(e)), 9, Xu)) : q("", !0),
-					G("div", Zu, [
-						e.language ? (U(), W("span", Qu, k(e.language), 1)) : q("", !0),
-						e.trained_at ? (U(), W("span", $u, k(I(uc)(e.trained_at)), 1)) : q("", !0),
-						e.recall === void 0 ? q("", !0) : (U(), W("span", ed, "recall " + k(e.recall), 1))
+					}, "Model · " + k(E(e)), 9, Qu)) : q("", !0),
+					G("div", $u, [
+						e.language ? (U(), W("span", ed, k(e.language), 1)) : q("", !0),
+						e.trained_at ? (U(), W("span", td, k(I(uc)(e.trained_at)), 1)) : q("", !0),
+						e.recall === void 0 ? q("", !0) : (U(), W("span", nd, "recall " + k(e.recall), 1))
 					])
 				]), G("button", {
 					type: "button",
 					disabled: !T(e),
 					onClick: (t) => I(ac)(T(e))
-				}, "Copy URL", 8, td)]))), 128))])) : (U(), W("div", Ku, "Train a wake word and its package will appear here."))])
+				}, "Copy URL", 8, rd)]))), 128))])) : (U(), W("div", Ju, "Train a wake word and its package will appear here."))]),
+				d[110] ||= G("div", { class: "native-notice esphome-notice" }, [G("strong", null, "ESPHome"), G("span", null, "Strict micro_wake_word manifest without Tater Native or calibration extensions.")], -1),
+				G("section", id, [d[108] ||= G("header", { class: "panel-head" }, [G("div", { class: "number" }, "ESP"), G("div", null, [G("h3", null, "ESPHome JSON"), G("p", null, "Use this URL as the model in an ESPHome micro_wake_word configuration.")])], -1), I(X).wakeWords.length ? (U(), W("div", od, [(U(!0), W(V, null, Lr(I(X).wakeWords, (e) => (U(), W("article", { key: `esphome-${e.key || re(e)}` }, [G("div", null, [
+					G("strong", null, k(e.label || e.name || "Trained wake word"), 1),
+					re(e) ? (U(), W("a", {
+						key: 0,
+						href: re(e),
+						target: "_blank",
+						rel: "noreferrer"
+					}, "ESPHome JSON · " + k(re(e)), 9, sd)) : (U(), W("span", cd, "ESPHome package URL unavailable")),
+					d[107] ||= G("div", { class: "meta-row" }, [G("span", null, "Schema v2"), G("span", null, "Same TFLite model")], -1)
+				]), G("button", {
+					type: "button",
+					disabled: !re(e),
+					onClick: (t) => I(ac)(re(e))
+				}, "Copy ESPHome URL", 8, ld)]))), 128))])) : (U(), W("div", ad, "ESPHome links appear after a wake word is trained."))])
 			], 64)) : q("", !0)], 64)) : (U(), W("div", Lc, [...d[46] ||= [G("span", { class: "spinner" }, null, -1), G("strong", null, "Connecting to the local trainer…", -1)]]))]),
 			(U(), ra(Jn, { to: "body" }, [I(X).consoleOpen ? (U(), W("div", {
 				key: 0,
 				class: "modal-backdrop console-backdrop",
 				onClick: d[39] ||= as((e) => I(X).consoleOpen = !1, ["self"])
-			}, [G("section", nd, [G("header", rd, [d[106] ||= G("div", null, [
+			}, [G("section", ud, [G("header", dd, [d[111] ||= G("div", null, [
 				G("span", { class: "eyebrow" }, "Live pipeline"),
 				G("h2", null, "Training Console"),
 				G("p", null, "Closing this window does not interrupt training.")
-			], -1), G("div", id, [
+			], -1), G("div", fd, [
 				r.value ? q("", !0) : (U(), W("button", {
 					key: 0,
 					type: "button",
@@ -4692,29 +4727,29 @@ var hc = {
 				onScrollPassive: v
 			}, [(U(!0), W(V, null, Lr(h.value, (e, t) => (U(), W("span", {
 				key: `${t}-${e}`,
-				class: O(E(e))
+				class: O(ie(e))
 			}, k(e), 3))), 128))], 544)])])) : q("", !0)])),
 			(U(), ra(Jn, { to: "body" }, [I(X).taterLinkOpen ? (U(), W("div", {
 				key: 0,
 				class: "modal-backdrop",
 				onClick: d[43] ||= as((e) => I(X).taterLinkOpen = !1, ["self"])
-			}, [G("section", ad, [G("header", od, [G("div", null, [
-				d[107] ||= G("span", { class: "eyebrow" }, "Secure pairing", -1),
+			}, [G("section", pd, [G("header", md, [G("div", null, [
+				d[112] ||= G("span", { class: "eyebrow" }, "Secure pairing", -1),
 				G("h2", null, k(o.value ? "Tater linked" : "Link Tater"), 1),
 				G("p", null, k(o.value ? "This trainer can securely publish wake-word updates." : "Enter the short-lived code shown in Tater Voice Settings."), 1)
 			]), G("button", {
 				type: "button",
 				onClick: d[40] ||= (e) => I(X).taterLinkOpen = !1
-			}, "Close")]), o.value ? (U(), W("div", sd, [
-				d[108] ||= G("i", null, "✓", -1),
+			}, "Close")]), o.value ? (U(), W("div", hd, [
+				d[113] ||= G("i", null, "✓", -1),
 				G("strong", null, "Successfully linked" + k(I(X).auto.trainer_link?.tater_name ? ` to ${I(X).auto.trainer_link.tater_name}` : ""), 1),
-				d[109] ||= G("span", null, "The private link key is stored locally and is never displayed.", -1)
-			])) : (U(), W("div", cd, [
-				G("label", ld, [d[110] ||= G("span", null, "Tater address", -1), R(G("input", {
+				d[114] ||= G("span", null, "The private link key is stored locally and is never displayed.", -1)
+			])) : (U(), W("div", gd, [
+				G("label", _d, [d[115] ||= G("span", null, "Tater address", -1), R(G("input", {
 					"onUpdate:modelValue": d[41] ||= (e) => i.value = e,
 					type: "text"
 				}, null, 512), [[Xo, i.value]])]),
-				G("label", ud, [d[111] ||= G("span", null, "Tater pairing code", -1), R(G("input", {
+				G("label", vd, [d[116] ||= G("span", null, "Tater pairing code", -1), R(G("input", {
 					id: "pairing-code",
 					"onUpdate:modelValue": d[42] ||= (e) => a.value = e,
 					class: "pairing-code",
@@ -4723,13 +4758,13 @@ var hc = {
 					autocomplete: "off",
 					onInput: w
 				}, null, 544), [[Xo, a.value]])]),
-				d[112] ||= G("small", null, "In Tater, open Voice Settings → Wake Word Trainer → Link Trainer.", -1),
+				d[117] ||= G("small", null, "In Tater, open Voice Settings → Wake Word Trainer → Link Trainer.", -1),
 				G("button", {
 					type: "button",
 					class: "button primary",
 					disabled: I(Z)("link"),
 					onClick: ee
-				}, k(I(Z)("link") ? "Linking securely…" : "Link Tater"), 9, dd)
+				}, k(I(Z)("link") ? "Linking securely…" : "Link Tater"), 9, yd)
 			]))])])) : q("", !0)])),
 			K(Ec),
 			K($a, { name: "toast" }, {
@@ -4742,7 +4777,7 @@ var hc = {
 			})
 		]));
 	}
-}), hd = document.getElementById("trainer-app");
-if (!hd) throw Error("Missing #trainer-app mount point");
-ds(md).mount(hd);
+}), Cd = document.getElementById("trainer-app");
+if (!Cd) throw Error("Missing #trainer-app mount point");
+ds(Sd).mount(Cd);
 //#endregion

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 +71,22 @@ class ModernTtsTests(unittest.TestCase):
             ["--position_temperature", "5.0", "--class_temperature", "0.0"],
         )
 
+    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 +604,87 @@ 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_session_stop.py

@@ -1,7 +1,11 @@
 from __future__ import annotations
 
+import io
 import signal
+import tempfile
+import threading
 import unittest
+from pathlib import Path
 from unittest.mock import patch
 
 import trainer_server as trainer
@@ -26,6 +30,19 @@ class _FakeTrainingProcess:
         self.returncode = -signal.SIGKILL
 
 
+class _CompletedTrainingProcess:
+    def __init__(self):
+        self.pid = 6543
+        self.returncode = 0
+        self.stdout = io.StringIO("worker started\n")
+
+    def poll(self):
+        return self.returncode
+
+    def wait(self, timeout=None):
+        return self.returncode
+
+
 class SessionStopTests(unittest.TestCase):
     def tearDown(self):
         trainer.TRAINING_STOP_EVENT.clear()
@@ -49,6 +66,51 @@ class SessionStopTests(unittest.TestCase):
             trainer.TRAINING_PROCESS = original_process
             trainer.TRAINING_THREAD = original_thread
 
+    def test_reserved_running_state_starts_the_background_worker(self):
+        original_process = trainer.TRAINING_PROCESS
+        original_thread = trainer.TRAINING_THREAD
+        original_raw_phrase = trainer.STATE.get("raw_phrase")
+        original_training = dict(trainer.STATE["training"])
+        try:
+            with tempfile.TemporaryDirectory() as directory:
+                data_dir = Path(directory)
+                process = _CompletedTrainingProcess()
+                trainer.TRAINING_PROCESS = None
+                trainer.TRAINING_THREAD = threading.current_thread()
+                with trainer.STATE_LOCK:
+                    trainer.STATE["raw_phrase"] = "hey tater"
+                    trainer.STATE["training"]["running"] = True
+
+                with (
+                    patch.object(trainer, "DATA_DIR", data_dir),
+                    patch.object(trainer, "_ensure_training_venv"),
+                    patch.object(trainer, "_ensure_training_datasets"),
+                    patch.object(trainer.subprocess, "Popen", return_value=process) as popen,
+                    patch.object(trainer, "_normalize_output_artifacts"),
+                ):
+                    trainer._run_training_background(
+                        "hey_tater",
+                        "en",
+                        True,
+                        auto_run=False,
+                        tts_mode="modern",
+                    )
+
+                popen.assert_called_once()
+                log_text = (data_dir / "recorder_training.log").read_text(encoding="utf-8")
+                self.assertIn("Nvidia Docker Training Run", log_text)
+                self.assertIn("worker started", log_text)
+                self.assertFalse(trainer.STATE["training"]["running"])
+                self.assertEqual(trainer.STATE["training"]["exit_code"], 0)
+                self.assertIsNone(trainer.TRAINING_THREAD)
+        finally:
+            with trainer.STATE_LOCK:
+                trainer.STATE["raw_phrase"] = original_raw_phrase
+                trainer.STATE["training"].clear()
+                trainer.STATE["training"].update(original_training)
+            trainer.TRAINING_PROCESS = original_process
+            trainer.TRAINING_THREAD = original_thread
+
 
 if __name__ == "__main__":
     unittest.main()

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:

trainer_server.py

@@ -581,6 +581,24 @@ 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 +637,11 @@ 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 +654,7 @@ 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,
@@ -2690,6 +2711,9 @@ 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 "",
@@ -3110,8 +3134,9 @@ def _run_training_background(
 
     with DATA_MANAGEMENT_LOCK:
         with STATE_LOCK:
-            if STATE["training"]["running"]:
-                return JSONResponse({"ok": False, "error": "Training already running"}, status_code=400)
+            # The API or auto-training scheduler reserves the run by setting
+            # this flag before the thread starts. Duplicate starts are already
+            # rejected there and by _start_training_thread's runtime lock.
             STATE["training"]["running"] = True
         STATE["training"]["exit_code"] = None
         STATE["training"]["log_lines"] = []
@@ -3959,6 +3984,24 @@ 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)