Release NVIDIA WakeWord Trainer v21

VERSION

@@ -1 +1 @@
-19
+21

WHATS_NEW.md

@@ -1,3 +1,4 @@
-- Fixed first-run Parakeet ONNX setup failing when its empty model directory was mistaken for a complete offline model.
+- Improved automatic review accuracy for short wake phrases that STT initially hears as similar-sounding words.
-- Parakeet now downloads or resumes the required INT8 snapshot before loading through ONNX ASR.
+- Added a conservative Faster Whisper confirmation pass that uses the currently configured wake phrase only when the unbiased transcript is already phonetically close.
-- Complete local snapshots are reused without Hub access, preserving offline startup after the initial download.
+- 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.

static/index.html

@@ -2341,6 +2341,7 @@
       if (item.average_probability !== null && item.average_probability !== undefined) meta.push(`<span class="pill">avg ${escapeHtml(item.average_probability)}</span>`);
       if (item.detection_profile) meta.push(`<span class="pill">profile ${escapeHtml(formatDetectionProfile(item.detection_profile))}</span>`);
       if (item.auto_review_status) meta.push(`<span class="pill ${item.auto_review_status === "error" ? "err" : "warn"}">auto ${escapeHtml(String(item.auto_review_status).replaceAll("_", " "))}</span>`);
+      if (item.auto_review_match_method === "guided_close_match") meta.push(`<span class="pill ok">guided wake confirmation</span>`);
       if (item.peak_probability_cutoff !== null && item.peak_probability_cutoff !== undefined) meta.push(`<span class="pill">peak cutoff ${escapeHtml(item.peak_probability_cutoff)}</span>`);
       if (item.probability_cutoff !== null && item.probability_cutoff !== undefined) meta.push(`<span class="pill">avg cutoff ${escapeHtml(item.probability_cutoff)}</span>`);
       if (item.active_window_count !== null && item.active_window_count !== undefined && item.min_active_windows !== null && item.min_active_windows !== undefined) {
@@ -2369,6 +2370,8 @@
           </div>
           <div class="fileMeta">${meta.join("") || `<span class="muted">No metadata attached</span>`}</div>
           ${item.transcript ? `<div class="autoAudit"><strong>STT transcript</strong><br>${escapeHtml(item.transcript)}</div>` : ""}
+          ${item.auto_review_guided_transcript ? `<div class="autoAudit"><strong>Guided wake check</strong><br>${escapeHtml(item.auto_review_guided_transcript)}</div>` : ""}
+          ${item.auto_review_reason ? `<div class="muted">${escapeHtml(item.auto_review_reason)}</div>` : ""}
           ${item.auto_review_error ? `<div class="muted">Auto review error: ${escapeHtml(item.auto_review_error)}</div>` : ""}
           <audio class="audioPlayer" controls preload="none" src="${escapeHtml(item.audio_url || `/api/audio/captured/${encodeURIComponent(item.saved_as)}`)}"></audio>
           <div class="muted">Stored as ${escapeHtml(item.saved_as)} · ${escapeHtml(formatSummary)}</div>

tests/test_auto_train.py

@@ -111,6 +111,21 @@ class AutoTrainTests(unittest.TestCase):
         self.assertTrue(trainer._transcript_contains_wake_phrase("Okay, HEY TATER!", "hey_tater"))
         self.assertFalse(trainer._transcript_contains_wake_phrase("Turn on the television", "hey tater"))
 
+    def test_phrase_similarity_recognizes_real_short_clip_mishearings(self):
+        for transcript in ("Hey, haters.", "Hate hater.", "Hey Ganger.", "Hey, gator."):
+            with self.subTest(transcript=transcript):
+                self.assertGreaterEqual(
+                    trainer._wake_phrase_similarity(transcript, "hey tater"),
+                    trainer.WAKE_PHRASE_GUIDANCE_MIN_SIMILARITY,
+                )
+
+        for transcript in ("turn on the lights", "what is the weather", "play some music"):
+            with self.subTest(transcript=transcript):
+                self.assertLess(
+                    trainer._wake_phrase_similarity(transcript, "hey tater"),
+                    trainer.WAKE_PHRASE_GUIDANCE_MIN_SIMILARITY,
+                )
+
     def test_stt_engine_selection_uses_managed_models(self):
         config = trainer._normalize_auto_train_config(
             {
@@ -160,6 +175,39 @@ class AutoTrainTests(unittest.TestCase):
         faster.assert_called_once()
         parakeet.assert_called_once()
 
+    def test_guided_faster_whisper_uses_dynamic_wake_phrase(self):
+        fake_model = SimpleNamespace(
+            transcribe=Mock(
+                return_value=(
+                    iter([SimpleNamespace(text=" hello "), SimpleNamespace(text="potato ")]),
+                    SimpleNamespace(),
+                )
+            )
+        )
+        with (
+            patch.object(
+                trainer,
+                "_resolve_faster_whisper_runtime",
+                return_value=("cuda", "float16"),
+            ),
+            patch.object(trainer, "_load_faster_whisper_model", return_value=fake_model),
+        ):
+            transcript = trainer._transcribe_capture_with_faster_whisper_guided(
+                Path("wake.wav"),
+                model="small.en",
+                language="en",
+                wake_phrase="Hello_Potato",
+            )
+
+        self.assertEqual(transcript, "hello potato")
+        _, kwargs = fake_model.transcribe.call_args
+        self.assertEqual(kwargs["hotwords"], "hello potato")
+        self.assertIn("hello potato", kwargs["initial_prompt"])
+        self.assertEqual(kwargs["beam_size"], 5)
+        self.assertEqual(kwargs["best_of"], 5)
+        self.assertEqual(kwargs["temperature"], 0.0)
+        self.assertFalse(kwargs["condition_on_previous_text"])
+
     def test_parakeet_loader_prefers_cuda_then_cpu(self):
         fake_model = object()
         fake_onnx_asr = SimpleNamespace(load_model=Mock(return_value=fake_model))
@@ -251,6 +299,7 @@ class AutoTrainTests(unittest.TestCase):
         self.assertNotIn('id="autoSttModel"', source)
         self.assertNotIn('id="autoSttDevice"', source)
         self.assertNotIn('id="autoSttComputeType"', source)
+        self.assertIn("Guided wake check", source)
 
     def test_phrase_miss_moves_wake_trigger_to_negative_samples(self):
         self.add_capture()
@@ -279,6 +328,80 @@ class AutoTrainTests(unittest.TestCase):
         self.assertEqual(metadata["auto_review_status"], "wake_phrase_detected")
         self.assertEqual(trainer.AUTO_TRAIN_STATE["pending_negative_count"], 0)
 
+    def test_close_transcript_uses_guided_faster_whisper_confirmation(self):
+        audio_path = self.add_capture()
+        with (
+            patch.object(trainer, "_transcribe_capture", return_value="Hey, haters."),
+            patch.object(
+                trainer,
+                "_transcribe_capture_with_faster_whisper_guided",
+                return_value="Hey Tater",
+            ) as guided,
+        ):
+            trainer._auto_review_capture("wake.wav")
+
+        self.assertTrue(audio_path.exists())
+        self.assertFalse(list(trainer.NEGATIVE_DIR.glob("*.wav")))
+        metadata = trainer._load_sidecar_json(audio_path)
+        self.assertEqual(metadata["auto_review_status"], "wake_phrase_detected")
+        self.assertEqual(metadata["transcript"], "Hey, haters.")
+        self.assertEqual(metadata["auto_review_guided_transcript"], "Hey Tater")
+        self.assertEqual(metadata["auto_review_match_method"], "guided_close_match")
+        self.assertGreaterEqual(
+            metadata["auto_review_phrase_similarity"],
+            trainer.WAKE_PHRASE_GUIDANCE_MIN_SIMILARITY,
+        )
+        guided.assert_called_once()
+        guided_args, guided_kwargs = guided.call_args
+        self.assertEqual(guided_args[0].resolve(), audio_path.resolve())
+        self.assertEqual(
+            guided_kwargs,
+            {
+                "model": "small.en",
+                "language": "en",
+                "wake_phrase": "hey tater",
+            },
+        )
+
+    def test_unconfirmed_close_transcript_stays_for_manual_review(self):
+        audio_path = self.add_capture()
+        with (
+            patch.object(trainer, "_transcribe_capture", return_value="Hate hater."),
+            patch.object(
+                trainer,
+                "_transcribe_capture_with_faster_whisper_guided",
+                return_value="Hate hater.",
+            ),
+        ):
+            trainer._auto_review_capture("wake.wav")
+
+        self.assertTrue(audio_path.exists())
+        self.assertFalse(list(trainer.NEGATIVE_DIR.glob("*.wav")))
+        metadata = trainer._load_sidecar_json(audio_path)
+        self.assertEqual(metadata["auto_review_status"], "wake_phrase_ambiguous")
+        self.assertEqual(metadata["transcript"], "Hate hater.")
+        self.assertEqual(metadata["auto_review_guided_transcript"], "Hate hater.")
+        self.assertEqual(trainer.AUTO_TRAIN_STATE["pending_negative_count"], 0)
+
+        self.assertEqual(trainer._queue_pending_auto_reviews(), 0)
+        self.assertEqual(trainer._queue_pending_auto_reviews(force=True), 1)
+
+    def test_close_parakeet_transcript_stays_for_manual_review(self):
+        audio_path = self.add_capture()
+        trainer.AUTO_TRAIN_CONFIG["stt_engine"] = trainer.STT_ENGINE_PARAKEET_ONNX
+        with (
+            patch.object(trainer, "_transcribe_capture", return_value="Hey Ganger."),
+            patch.object(trainer, "_transcribe_capture_with_faster_whisper_guided") as guided,
+        ):
+            trainer._auto_review_capture("wake.wav")
+
+        guided.assert_not_called()
+        self.assertTrue(audio_path.exists())
+        self.assertFalse(list(trainer.NEGATIVE_DIR.glob("*.wav")))
+        metadata = trainer._load_sidecar_json(audio_path)
+        self.assertEqual(metadata["auto_review_status"], "wake_phrase_ambiguous")
+        self.assertEqual(metadata["auto_review_stt_engine"], "parakeet_onnx")
+
     def test_matching_phrase_is_deleted_when_cleanup_is_enabled(self):
         audio_path = self.add_capture()
         trainer.AUTO_TRAIN_CONFIG["delete_confirmed_wakes"] = True

trainer_server.py

@@ -20,6 +20,7 @@ import unicodedata
 import wave
 from array import array
 from datetime import datetime, timedelta, timezone
+from difflib import SequenceMatcher
 from math import isfinite, log10
 from pathlib import Path
 from typing import Dict, Any, List, Callable, Optional, Tuple
@@ -118,6 +119,7 @@ DEFAULT_PARAKEET_ONNX_REPO = os.environ.get(
     "istupakov/parakeet-tdt-0.6b-v3-onnx",
 )
 DEFAULT_PARAKEET_ONNX_QUANTIZATION = "int8"
+WAKE_PHRASE_GUIDANCE_MIN_SIMILARITY = 0.68
 
 AUTO_TRAIN_DEFAULT_CONFIG: Dict[str, Any] = {
     "enabled": False,
@@ -835,6 +837,28 @@ def _transcript_contains_wake_phrase(transcript: Any, wake_phrase: Any) -> bool:
     return f" {normalized_phrase} " in f" {normalized_transcript} "
 
 
+def _wake_phrase_similarity(transcript: Any, wake_phrase: Any) -> float:
+    transcript_words = _normalize_transcript_text(transcript).split()
+    phrase_words = _normalize_transcript_text(wake_phrase).split()
+    if not transcript_words or not phrase_words:
+        return 0.0
+    if _transcript_contains_wake_phrase(transcript, wake_phrase):
+        return 1.0
+
+    phrase_token = "".join(phrase_words)
+    minimum_words = max(1, len(phrase_words) - 1)
+    maximum_words = min(len(transcript_words), len(phrase_words) + 1)
+    best_score = 0.0
+    for word_count in range(minimum_words, maximum_words + 1):
+        for start in range(0, len(transcript_words) - word_count + 1):
+            candidate = "".join(transcript_words[start : start + word_count])
+            best_score = max(
+                best_score,
+                SequenceMatcher(None, candidate, phrase_token).ratio(),
+            )
+    return best_score
+
+
 def _captured_event_is_close_miss(metadata: Dict[str, Any]) -> bool:
     event_type = str(metadata.get("event_type") or "captured").strip().lower()
     return "close" in event_type
@@ -925,6 +949,40 @@ def _transcribe_capture_with_faster_whisper(audio_path: Path, *, model: str, lan
     return transcript
 
 
+def _transcribe_capture_with_faster_whisper_guided(
+    audio_path: Path,
+    *,
+    model: str,
+    language: str,
+    wake_phrase: str,
+) -> str:
+    normalized_phrase = _normalize_transcript_text(wake_phrase)
+    if not normalized_phrase:
+        return ""
+    device, compute_type = _resolve_faster_whisper_runtime("auto", "auto")
+    whisper_model = _load_faster_whisper_model(
+        model_name=model,
+        device=device,
+        compute_type=compute_type,
+    )
+    with FASTER_WHISPER_TRANSCRIBE_LOCK:
+        segments, _info = whisper_model.transcribe(
+            str(audio_path),
+            language=language or None,
+            beam_size=5,
+            best_of=5,
+            temperature=0.0,
+            condition_on_previous_text=False,
+            initial_prompt=f'The wake phrase is "{normalized_phrase}".',
+            hotwords=normalized_phrase,
+        )
+        return re.sub(
+            r"\s+",
+            " ",
+            " ".join(str(segment.text or "").strip() for segment in segments),
+        ).strip()
+
+
 def _parakeet_onnx_providers() -> List[str]:
     try:
         import onnxruntime as ort
@@ -957,11 +1015,11 @@ def _load_parakeet_onnx_model():
         cached = PARAKEET_ONNX_MODEL_CACHE.get(cache_key)
         if cached is not None:
             return cached
-    suffix = (
+        suffix = (
-        f".{DEFAULT_PARAKEET_ONNX_QUANTIZATION}"
+            f".{DEFAULT_PARAKEET_ONNX_QUANTIZATION}"
-        if DEFAULT_PARAKEET_ONNX_QUANTIZATION
+            if DEFAULT_PARAKEET_ONNX_QUANTIZATION
-        else ""
+            else ""
-    )
+        )
         model_patterns = [
             "config.json",
             "vocab.txt",
@@ -988,10 +1046,10 @@ def _load_parakeet_onnx_model():
         os.environ["HUGGINGFACE_HUB_CACHE"] = str(AUTO_TRAIN_MODEL_DIR / "hub")
         try:
             snapshot_root = AUTO_TRAIN_MODEL_DIR
-    if not all(
+            if not all(
-        (AUTO_TRAIN_MODEL_DIR / filename).is_file()
+                (AUTO_TRAIN_MODEL_DIR / filename).is_file()
-        for filename in required_model_files
+                for filename in required_model_files
-    ):
+            ):
                 from huggingface_hub import snapshot_download
 
                 snapshot_root = Path(
@@ -1119,7 +1177,7 @@ def _queue_pending_auto_reviews(*, force: bool = False) -> int:
             metadata.pop("auto_review_status", None)
             _write_sidecar_json(audio_path, metadata)
             status = ""
-        if force and status in {"error", "no_speech"}:
+        if force and status in {"error", "no_speech", "wake_phrase_ambiguous"}:
             metadata.pop("auto_review_status", None)
             _write_sidecar_json(audio_path, metadata)
             status = ""
@@ -1220,11 +1278,38 @@ def _auto_review_capture(file_name: str) -> None:
             _record_auto_review_result(file_name=file_name, transcript=transcript, result="no_speech")
             return
 
-        if _transcript_contains_wake_phrase(transcript, wake_phrase):
+        phrase_similarity = _wake_phrase_similarity(transcript, wake_phrase)
+        phrase_detected = _transcript_contains_wake_phrase(transcript, wake_phrase)
+        match_method = "exact" if phrase_detected else ""
+        metadata["auto_review_phrase_similarity"] = round(phrase_similarity, 4)
+
+        if (
+            not phrase_detected
+            and phrase_similarity >= WAKE_PHRASE_GUIDANCE_MIN_SIMILARITY
+            and stt_engine == STT_ENGINE_FASTER_WHISPER
+        ):
+            guided_transcript = _transcribe_capture_with_faster_whisper_guided(
+                audio_path,
+                model=str(metadata["auto_review_stt_model"]),
+                language=str(config.get("language") or DEFAULT_LANGUAGE),
+                wake_phrase=wake_phrase,
+            )
+            metadata["auto_review_guided_transcript"] = guided_transcript
+            if _transcript_contains_wake_phrase(guided_transcript, wake_phrase):
+                phrase_detected = True
+                match_method = "guided_close_match"
+
+        if match_method:
+            metadata["auto_review_match_method"] = match_method
+
+        if phrase_detected:
+            guided_confirmation = match_method == "guided_close_match"
             if is_close_miss:
                 metadata["auto_review_status"] = "approved_positive"
                 metadata["auto_review_reason"] = (
-                    "Close miss contained the configured wake phrase and was promoted to a positive sample."
+                    "Close miss was confirmed as the configured wake phrase and promoted to a positive sample."
+                    if guided_confirmation
+                    else "Close miss contained the configured wake phrase and was promoted to a positive sample."
                 )
                 metadata["auto_positive"] = True
                 _write_sidecar_json(audio_path, metadata)
@@ -1249,11 +1334,29 @@ def _auto_review_capture(file_name: str) -> None:
                 )
                 return
             metadata["auto_review_status"] = "wake_phrase_detected"
-            metadata["auto_review_reason"] = "Wake phrase found in transcript; left for manual positive review."
+            metadata["auto_review_reason"] = (
+                "Wake phrase confirmed by a guided second STT pass; left for manual positive review."
+                if guided_confirmation
+                else "Wake phrase found in transcript; left for manual positive review."
+            )
             _write_sidecar_json(audio_path, metadata)
             _record_auto_review_result(file_name=file_name, transcript=transcript, result="wake_phrase_detected")
             return
 
+        if phrase_similarity >= WAKE_PHRASE_GUIDANCE_MIN_SIMILARITY:
+            metadata["auto_review_status"] = "wake_phrase_ambiguous"
+            metadata["auto_review_reason"] = (
+                "STT sounded close to the configured wake phrase but could not confirm it; "
+                "left for manual review."
+            )
+            _write_sidecar_json(audio_path, metadata)
+            _record_auto_review_result(
+                file_name=file_name,
+                transcript=transcript,
+                result="wake_phrase_ambiguous",
+            )
+            return
+
         if is_close_miss:
             metadata["auto_review_status"] = "close_miss_phrase_not_detected"
             metadata["auto_review_reason"] = (
@@ -2162,6 +2265,9 @@ def _captured_item_from_path(audio_path: Path) -> Dict[str, Any]:
         "auto_review_status": meta.get("auto_review_status") or "",
         "auto_review_reason": meta.get("auto_review_reason") or "",
         "auto_review_error": meta.get("auto_review_error") or "",
+        "auto_review_guided_transcript": meta.get("auto_review_guided_transcript") or "",
+        "auto_review_phrase_similarity": meta.get("auto_review_phrase_similarity"),
+        "auto_review_match_method": meta.get("auto_review_match_method") or "",
         "size_bytes": stat.st_size,
         "audio_url": f"/api/audio/captured/{audio_path.name}",
     }