diff --git a/README.md b/README.md
index 553b4ec..d432353 100644
--- a/README.md
+++ b/README.md
@@ -171,7 +171,7 @@ Starting a new session does not clear samples. Use the clear buttons in `Samples
For each new wake-trigger clip sent to the trainer:
-1. Faster Whisper transcribes the audio locally.
+1. The selected local STT engine transcribes the audio.
2. If the transcript contains the configured wake phrase, the clip stays in `Captured Audio` for manual review by default.
3. If speech was transcribed but the wake phrase is absent, the clip moves to `/data/negative_samples/` as an auto-reviewed hard negative.
4. Empty transcripts, VAD-blocked captures, and captures for another wake word stay out of the automatic negative path.
@@ -183,7 +183,7 @@ Two optional cleanup rules are available:
A close miss with an empty transcript or without the configured phrase stays in `Captured Audio`; it is never turned into a negative automatically. Saving Auto Training settings also scans existing eligible captures. Enabling close-miss promotion reviews previous unreviewed close misses, while enabling cleanup removes previously confirmed good wakes without transcribing them a second time.
-The default `small.en` model uses CUDA with `float16` when CTranslate2 can see an NVIDIA GPU, and falls back to CPU with `int8`. Choose a multilingual Faster Whisper model such as `small` when the wake phrase is not English. Downloaded STT models are cached in `/data/auto_train_models/`.
+Auto Training exposes only an engine selector. Faster Whisper is the recommended default and uses CUDA with `float16` when CTranslate2 can see an NVIDIA GPU, with a CPU `int8` fallback. The trainer manages `small.en` for English and `small` for other languages. Parakeet ONNX uses the managed INT8 `nemo-parakeet-tdt-0.6b-v3` model with CUDA and CPU fallback. Downloaded STT models are cached in `/data/auto_train_models/`.
Scheduled training runs only after the configured number of new automatic negatives has accumulated. A successful run securely publishes the trained wake-word name and JSON URL to the linked Tater instance. Tater saves it as the global satellite wake word and pushes the updated setting to every connected satellite, so no satellite firmware change is required.
diff --git a/VERSION b/VERSION
index 98d9bcb..3c03207 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-17
+18
diff --git a/WHATS_NEW.md b/WHATS_NEW.md
index ebf2aa8..cde4096 100644
--- a/WHATS_NEW.md
+++ b/WHATS_NEW.md
@@ -1,2 +1,3 @@
-- Fixed the training-status endpoint crashing after a training log was created because its log-tail limits were missing.
-- Restored bounded, incremental training-log updates so the UI can continue showing live progress without repeatedly reading the entire log.
+- Added Parakeet ONNX as a second local Auto Training STT engine alongside Faster Whisper.
+- Replaced manual model, device, and compute fields with a simple engine selector and managed language-aware models.
+- Added CUDA-enabled ONNX Runtime with CPU fallback, runtime reporting, and model-cache cleanup when switching engines.
diff --git a/run.sh b/run.sh
index 53ef07f..1d61a41 100644
--- a/run.sh
+++ b/run.sh
@@ -33,8 +33,11 @@ install_ui_deps() {
"silero-vad>=5.0.0" \
"numpy>=1.24.0" \
"faster-whisper>=1.0.0" \
+ "onnx-asr[hub]>=0.12.0" \
"nvidia-cublas-cu12" \
"nvidia-cudnn-cu12==9.*"
+ ${PIP} uninstall -y onnxruntime
+ ${PIP} install "onnxruntime-gpu[cuda,cudnn]<1.27"
}
# -----------------------------
@@ -82,11 +85,13 @@ minimum = {
"silero-vad": "5.0.0",
"numpy": "1.24.0",
"faster-whisper": "1.0.0",
+ "onnx-asr": "0.12.0",
"nvidia-cudnn-cu12": "9.0.0",
}
present = (
"torch",
"nvidia-cublas-cu12",
+ "onnxruntime-gpu",
)
for package, expected in exact.items():
@@ -97,6 +102,10 @@ for package, minimum_version in minimum.items():
raise SystemExit(1)
for package in present:
md.version(package)
+
+import onnxruntime as ort
+if "CUDAExecutionProvider" not in ort.get_available_providers():
+ raise SystemExit(1)
PY
then
echo "UI dependencies missing or stale; installing recorder dependencies"
diff --git a/static/index.html b/static/index.html
index fb69d6d..d6c08e4 100644
--- a/static/index.html
+++ b/static/index.html
@@ -1348,7 +1348,7 @@
Enable Auto Training Eligible wake triggers will be queued for local Faster Whisper transcription.Enable Auto Training Eligible wake triggers will be queued for transcription with the selected local STT engine.
- Faster Whisper model
-
-
- STT device
-
- Auto (prefer CUDA)
- CUDA
- CPU
-
-
-
- Compute type
-
- Auto (float16 CUDA / int8 CPU)
- float16
- int8_float16
- int8
- float32
- CTranslate2 default
+ STT engine
+
+ Faster Whisper (recommended)
+ Parakeet ONNX
@@ -2091,9 +2075,7 @@
$("autoEnabled").checked = Boolean(config.enabled);
$("autoWakePhrase").value = config.wake_phrase || uiState.session?.raw_phrase || "";
$("autoLanguage").value = config.language || uiState.session?.language || "en";
- $("autoSttModel").value = config.stt_model || "small.en";
- $("autoSttDevice").value = config.stt_device || "auto";
- $("autoSttComputeType").value = config.stt_compute_type || "auto";
+ $("autoSttEngine").value = config.stt_engine || "faster_whisper";
$("autoMinimumChars").value = String(config.minimum_transcript_chars ?? 2);
$("autoDeleteConfirmedWakes").checked = Boolean(config.delete_confirmed_wakes);
$("autoPromoteCloseMisses").checked = Boolean(config.promote_close_misses);
@@ -2138,7 +2120,10 @@
if (state.last_review_file) audit.push(state.last_review_file);
if (state.last_review_transcript) audit.push(`STT: “${state.last_review_transcript}”`);
if (state.last_review_error) audit.push(`Error: ${state.last_review_error}`);
- if (state.last_stt_device) audit.push(`STT runtime: ${state.last_stt_device} / ${state.last_stt_compute_type || "default"}`);
+ if (state.last_stt_engine) {
+ const runtimeLabel = [state.last_stt_device, state.last_stt_compute_type].filter(Boolean).join(" / ");
+ audit.push(`STT engine: ${String(state.last_stt_engine).replaceAll("_", " ")}${runtimeLabel ? ` · ${runtimeLabel}` : ""}`);
+ }
if (state.last_notify_at) {
audit.push(state.last_notify_error
? `Wake-word publish failed: ${state.last_notify_error}`
@@ -2159,9 +2144,7 @@
enabled: $("autoEnabled").checked,
wake_phrase: ($("autoWakePhrase").value || "").trim(),
language: ($("autoLanguage").value || "en").trim(),
- stt_model: ($("autoSttModel").value || "").trim(),
- stt_device: $("autoSttDevice").value || "auto",
- stt_compute_type: $("autoSttComputeType").value || "auto",
+ stt_engine: $("autoSttEngine").value || "faster_whisper",
minimum_transcript_chars: Number($("autoMinimumChars").value || 2),
delete_confirmed_wakes: $("autoDeleteConfirmedWakes").checked,
promote_close_misses: $("autoPromoteCloseMisses").checked,
diff --git a/tests/test_auto_train.py b/tests/test_auto_train.py
index a18edd1..4ef9b5d 100644
--- a/tests/test_auto_train.py
+++ b/tests/test_auto_train.py
@@ -62,8 +62,6 @@ class AutoTrainTests(unittest.TestCase):
"wake_phrase": "hey tater",
"language": "en",
"tater_url": "http://127.0.0.1:8501",
- "stt_device": "auto",
- "stt_compute_type": "auto",
}
)
)
@@ -110,9 +108,90 @@ 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_stt_engine_selection_uses_managed_models(self):
+ config = trainer._normalize_auto_train_config(
+ {
+ "stt_engine": "parakeet-onnx",
+ "stt_model": "user/should-not-be-used",
+ "stt_device": "cpu",
+ "stt_compute_type": "float32",
+ }
+ )
+
+ self.assertEqual(config["stt_engine"], trainer.STT_ENGINE_PARAKEET_ONNX)
+ self.assertNotIn("stt_model", config)
+ self.assertNotIn("stt_device", config)
+ self.assertNotIn("stt_compute_type", config)
+ self.assertEqual(
+ trainer._managed_stt_model(config["stt_engine"], "en"),
+ trainer.DEFAULT_PARAKEET_ONNX_MODEL,
+ )
+ self.assertEqual(
+ trainer._managed_stt_model(trainer.STT_ENGINE_FASTER_WHISPER, "de"),
+ trainer.DEFAULT_FASTER_WHISPER_MULTILINGUAL_MODEL,
+ )
+
+ def test_stt_router_supports_both_nvidia_engines(self):
+ audio_path = Path("wake.wav")
+ with (
+ patch.object(trainer, "_transcribe_capture_with_faster_whisper", return_value="faster") as faster,
+ patch.object(trainer, "_transcribe_capture_with_parakeet", return_value="parakeet") as parakeet,
+ ):
+ self.assertEqual(
+ trainer._transcribe_capture(
+ audio_path,
+ engine=trainer.STT_ENGINE_FASTER_WHISPER,
+ language="en",
+ ),
+ "faster",
+ )
+ self.assertEqual(
+ trainer._transcribe_capture(
+ audio_path,
+ engine=trainer.STT_ENGINE_PARAKEET_ONNX,
+ language="en",
+ ),
+ "parakeet",
+ )
+
+ faster.assert_called_once()
+ parakeet.assert_called_once()
+
+ def test_parakeet_loader_prefers_cuda_then_cpu(self):
+ fake_model = object()
+ fake_onnx_asr = SimpleNamespace(load_model=Mock(return_value=fake_model))
+ with (
+ patch.dict(sys.modules, {"onnx_asr": fake_onnx_asr}),
+ patch.object(
+ trainer,
+ "_parakeet_onnx_providers",
+ return_value=["CUDAExecutionProvider", "CPUExecutionProvider"],
+ ),
+ ):
+ with trainer.PARAKEET_ONNX_MODEL_LOCK:
+ trainer.PARAKEET_ONNX_MODEL_CACHE.clear()
+ loaded = trainer._load_parakeet_onnx_model()
+
+ self.assertIs(loaded, fake_model)
+ fake_onnx_asr.load_model.assert_called_once_with(
+ trainer.DEFAULT_PARAKEET_ONNX_MODEL,
+ str(trainer.AUTO_TRAIN_MODEL_DIR),
+ quantization="int8",
+ providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+ )
+
+ def test_ui_exposes_engine_selector_without_manual_runtime_fields(self):
+ source = (Path(__file__).resolve().parents[1] / "static" / "index.html").read_text(
+ encoding="utf-8"
+ )
+ self.assertIn('id="autoSttEngine"', source)
+ self.assertNotIn('id="autoSttModel"', source)
+ self.assertNotIn('id="autoSttDevice"', source)
+ self.assertNotIn('id="autoSttComputeType"', source)
+
def test_phrase_miss_moves_wake_trigger_to_negative_samples(self):
self.add_capture()
- with patch.object(trainer, "_transcribe_capture_with_faster_whisper", return_value="turn on the kitchen lights"):
+ with patch.object(trainer, "_transcribe_capture", return_value="turn on the kitchen lights"):
trainer._auto_review_capture("wake.wav")
self.assertFalse((trainer.CAPTURED_DIR / "wake.wav").exists())
@@ -122,11 +201,13 @@ class AutoTrainTests(unittest.TestCase):
self.assertTrue(metadata["auto_negative"])
self.assertEqual(metadata["review_status"], "auto_approved_negative")
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")
self.assertEqual(trainer.AUTO_TRAIN_STATE["pending_negative_count"], 1)
def test_matching_phrase_stays_in_manual_review_inbox(self):
audio_path = self.add_capture()
- with patch.object(trainer, "_transcribe_capture_with_faster_whisper", return_value="hey tater turn on the lights"):
+ with patch.object(trainer, "_transcribe_capture", return_value="hey tater turn on the lights"):
trainer._auto_review_capture("wake.wav")
self.assertTrue(audio_path.exists())
@@ -140,7 +221,7 @@ class AutoTrainTests(unittest.TestCase):
trainer.AUTO_TRAIN_CONFIG["delete_confirmed_wakes"] = True
with patch.object(
trainer,
- "_transcribe_capture_with_faster_whisper",
+ "_transcribe_capture",
return_value="hey tater turn on the lights",
):
trainer._auto_review_capture("wake.wav")
@@ -164,7 +245,7 @@ class AutoTrainTests(unittest.TestCase):
trainer.AUTO_TRAIN_CONFIG["delete_confirmed_wakes"] = True
self.assertEqual(trainer._queue_pending_auto_reviews(), 1)
- with patch.object(trainer, "_transcribe_capture_with_faster_whisper") as transcribe:
+ with patch.object(trainer, "_transcribe_capture") as transcribe:
trainer._auto_review_capture("wake.wav")
transcribe.assert_not_called()
@@ -173,7 +254,7 @@ class AutoTrainTests(unittest.TestCase):
def test_close_miss_is_not_transcribed_by_default(self):
audio_path = self.add_capture(event_type="close_miss")
- with patch.object(trainer, "_transcribe_capture_with_faster_whisper") as transcribe:
+ with patch.object(trainer, "_transcribe_capture") as transcribe:
trainer._auto_review_capture("wake.wav")
transcribe.assert_not_called()
@@ -190,7 +271,7 @@ class AutoTrainTests(unittest.TestCase):
def test_close_miss_with_phrase_is_promoted_when_enabled(self):
self.add_capture(event_type="close_miss")
trainer.AUTO_TRAIN_CONFIG["promote_close_misses"] = True
- with patch.object(trainer, "_transcribe_capture_with_faster_whisper", return_value="hey tater"):
+ with patch.object(trainer, "_transcribe_capture", return_value="hey tater"):
trainer._auto_review_capture("wake.wav")
self.assertFalse((trainer.CAPTURED_DIR / "wake.wav").exists())
@@ -208,7 +289,7 @@ class AutoTrainTests(unittest.TestCase):
trainer.AUTO_TRAIN_CONFIG["promote_close_misses"] = True
with patch.object(
trainer,
- "_transcribe_capture_with_faster_whisper",
+ "_transcribe_capture",
return_value="turn on the lights",
):
trainer._auto_review_capture("wake.wav")
@@ -222,7 +303,7 @@ class AutoTrainTests(unittest.TestCase):
def test_vad_blocked_close_miss_is_never_transcribed(self):
audio_path = self.add_capture(event_type="close_miss", blocked_by_vad=True)
trainer.AUTO_TRAIN_CONFIG["promote_close_misses"] = True
- with patch.object(trainer, "_transcribe_capture_with_faster_whisper") as transcribe:
+ with patch.object(trainer, "_transcribe_capture") as transcribe:
trainer._auto_review_capture("wake.wav")
transcribe.assert_not_called()
@@ -231,7 +312,7 @@ class AutoTrainTests(unittest.TestCase):
def test_capture_for_another_wake_word_is_not_transcribed(self):
audio_path = self.add_capture(wake_word="computer")
- with patch.object(trainer, "_transcribe_capture_with_faster_whisper") as transcribe:
+ with patch.object(trainer, "_transcribe_capture") as transcribe:
trainer._auto_review_capture("wake.wav")
transcribe.assert_not_called()
diff --git a/tests/test_run_sh.py b/tests/test_run_sh.py
index aafd83f..f4afa57 100644
--- a/tests/test_run_sh.py
+++ b/tests/test_run_sh.py
@@ -61,6 +61,13 @@ class RunShCudaLibraryPathTests(unittest.TestCase):
self.assertEqual(result.returncode, 0, result.stderr)
self.assertEqual(result.stdout.strip(), "")
+ def test_parakeet_uses_cuda_onnxruntime_package(self) -> None:
+ source = RUN_SH.read_text(encoding="utf-8")
+
+ self.assertIn('"onnx-asr[hub]>=0.12.0"', source)
+ self.assertIn('"onnxruntime-gpu[cuda,cudnn]<1.27"', source)
+ self.assertIn('"CUDAExecutionProvider"', source)
+
if __name__ == "__main__":
unittest.main()
diff --git a/trainer_server.py b/trainer_server.py
index fa1bafb..5fc47a8 100644
--- a/trainer_server.py
+++ b/trainer_server.py
@@ -2,6 +2,7 @@
# trainer_server.py
import contextlib
+import gc
import io
import os
import queue
@@ -88,15 +89,37 @@ TARGET_SAMPLE_RATE = 16000
TARGET_CHANNELS = 1
TARGET_SAMPLE_WIDTH_BYTES = 2
CAPTURE_GAIN_PROFILE = "capture_rms_v1"
-DEFAULT_FASTER_WHISPER_MODEL = os.environ.get("AUTO_TRAIN_STT_MODEL", "small.en")
+STT_ENGINE_FASTER_WHISPER = "faster_whisper"
+STT_ENGINE_PARAKEET_ONNX = "parakeet_onnx"
+SUPPORTED_STT_ENGINES = {
+ STT_ENGINE_FASTER_WHISPER,
+ STT_ENGINE_PARAKEET_ONNX,
+}
+DEFAULT_STT_ENGINE = os.environ.get(
+ "AUTO_TRAIN_STT_ENGINE",
+ STT_ENGINE_FASTER_WHISPER,
+).strip().lower().replace("-", "_")
+if DEFAULT_STT_ENGINE not in SUPPORTED_STT_ENGINES:
+ DEFAULT_STT_ENGINE = STT_ENGINE_FASTER_WHISPER
+DEFAULT_FASTER_WHISPER_EN_MODEL = os.environ.get(
+ "AUTO_TRAIN_FASTER_WHISPER_EN_MODEL",
+ "small.en",
+)
+DEFAULT_FASTER_WHISPER_MULTILINGUAL_MODEL = os.environ.get(
+ "AUTO_TRAIN_FASTER_WHISPER_MULTILINGUAL_MODEL",
+ "small",
+)
+DEFAULT_PARAKEET_ONNX_MODEL = os.environ.get(
+ "AUTO_TRAIN_PARAKEET_ONNX_MODEL",
+ "nemo-parakeet-tdt-0.6b-v3",
+)
+DEFAULT_PARAKEET_ONNX_QUANTIZATION = "int8"
AUTO_TRAIN_DEFAULT_CONFIG: Dict[str, Any] = {
"enabled": False,
"wake_phrase": "",
"language": DEFAULT_LANGUAGE,
- "stt_model": DEFAULT_FASTER_WHISPER_MODEL,
- "stt_device": "auto",
- "stt_compute_type": "auto",
+ "stt_engine": DEFAULT_STT_ENGINE,
"minimum_transcript_chars": 2,
"delete_confirmed_wakes": False,
"promote_close_misses": False,
@@ -119,6 +142,8 @@ AUTO_TRAIN_DEFAULT_STATE: Dict[str, Any] = {
"last_review_transcript": "",
"last_review_result": "",
"last_review_error": "",
+ "last_stt_engine": "",
+ "last_stt_model": "",
"last_stt_device": "",
"last_stt_compute_type": "",
"last_train_started_at": "",
@@ -186,6 +211,10 @@ AUTO_TRAIN_RUNTIME: Dict[str, Any] = {
LAN_ADDRESS_CACHE: Dict[str, Any] = {"value": "", "fetched_at": 0.0}
FASTER_WHISPER_MODEL_LOCK = threading.RLock()
FASTER_WHISPER_MODEL_CACHE: Dict[Tuple[str, str, str], Any] = {}
+FASTER_WHISPER_TRANSCRIBE_LOCK = threading.RLock()
+PARAKEET_ONNX_MODEL_LOCK = threading.RLock()
+PARAKEET_ONNX_MODEL_CACHE: Dict[Tuple[str, str, Tuple[str, ...]], Any] = {}
+PARAKEET_ONNX_TRANSCRIBE_LOCK = threading.RLock()
PIPER_CATALOG_CACHE: Dict[str, Any] = {
"fetched_at": 0.0,
"entries": None,
@@ -461,25 +490,60 @@ def _normalize_http_base_url(value: Any, *, allow_empty: bool = True) -> str:
return token
+def _normalize_stt_engine(value: Any) -> str:
+ token = str(value or DEFAULT_STT_ENGINE).strip().lower().replace("-", "_")
+ aliases = {
+ "faster": STT_ENGINE_FASTER_WHISPER,
+ "fasterwhisper": STT_ENGINE_FASTER_WHISPER,
+ "parakeet": STT_ENGINE_PARAKEET_ONNX,
+ "onnx_parakeet": STT_ENGINE_PARAKEET_ONNX,
+ }
+ token = aliases.get(token, token)
+ if token not in SUPPORTED_STT_ENGINES:
+ raise ValueError("STT engine must be Faster Whisper or Parakeet ONNX.")
+ return token
+
+
+def _managed_stt_model(engine: Any, language: Any = DEFAULT_LANGUAGE) -> str:
+ token = _normalize_stt_engine(engine)
+ language_token = str(language or DEFAULT_LANGUAGE).strip().lower().replace("-", "_")
+ english = language_token == "en" or language_token.startswith("en_")
+ if token == STT_ENGINE_FASTER_WHISPER:
+ return (
+ DEFAULT_FASTER_WHISPER_EN_MODEL
+ if english
+ else DEFAULT_FASTER_WHISPER_MULTILINGUAL_MODEL
+ )
+ return DEFAULT_PARAKEET_ONNX_MODEL
+
+
+def _stt_engine_catalog(language: Any = DEFAULT_LANGUAGE) -> List[Dict[str, Any]]:
+ return [
+ {
+ "value": STT_ENGINE_FASTER_WHISPER,
+ "label": "Faster Whisper",
+ "model": _managed_stt_model(STT_ENGINE_FASTER_WHISPER, language),
+ "recommended": True,
+ },
+ {
+ "value": STT_ENGINE_PARAKEET_ONNX,
+ "label": "Parakeet ONNX",
+ "model": _managed_stt_model(STT_ENGINE_PARAKEET_ONNX, language),
+ },
+ ]
+
+
def _normalize_auto_train_config(values: Dict[str, Any] | None, *, base: Dict[str, Any] | None = None) -> Dict[str, Any]:
incoming = values if isinstance(values, dict) else {}
source = {**AUTO_TRAIN_DEFAULT_CONFIG, **(base or {}), **incoming}
schedule_hours = _bounded_int(source.get("schedule_hours"), 24, 0, 24 * 30)
language = str(source.get("language") or DEFAULT_LANGUAGE).strip().lower().replace("-", "_")
language = re.sub(r"[^a-z0-9_]", "", language) or DEFAULT_LANGUAGE
- stt_device = str(source.get("stt_device") or "auto").strip().lower()
- if stt_device not in {"auto", "cuda", "cpu"}:
- raise ValueError("Faster Whisper device must be auto, cuda, or cpu.")
- stt_compute_type = str(source.get("stt_compute_type") or "auto").strip().lower()
- if stt_compute_type not in {"auto", "default", "float16", "float32", "int8", "int8_float16"}:
- raise ValueError("Unsupported Faster Whisper compute type.")
return {
"enabled": _config_bool(source.get("enabled")),
"wake_phrase": str(source.get("wake_phrase") or "").strip(),
"language": language,
- "stt_model": str(source.get("stt_model") or DEFAULT_FASTER_WHISPER_MODEL).strip() or DEFAULT_FASTER_WHISPER_MODEL,
- "stt_device": stt_device,
- "stt_compute_type": stt_compute_type,
+ "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")),
"promote_close_misses": _config_bool(source.get("promote_close_misses")),
@@ -555,10 +619,12 @@ def _public_auto_train_config() -> Dict[str, Any]:
def _auto_train_status_payload() -> Dict[str, Any]:
with AUTO_TRAIN_LOCK:
+ language = AUTO_TRAIN_CONFIG.get("language") or DEFAULT_LANGUAGE
return {
"config": _public_auto_train_config(),
"state": dict(AUTO_TRAIN_STATE),
"runtime": dict(AUTO_TRAIN_RUNTIME),
+ "stt_engines": _stt_engine_catalog(language),
"advertised_base_url": _advertised_base_url(),
"trainer_link": _tater_link_public_status(),
}
@@ -828,33 +894,163 @@ def _load_faster_whisper_model(*, model_name: str, device: str, compute_type: st
def _transcribe_capture_with_faster_whisper(audio_path: Path, *, model: str, language: str) -> str:
- with AUTO_TRAIN_LOCK:
- device_value = AUTO_TRAIN_CONFIG.get("stt_device")
- compute_value = AUTO_TRAIN_CONFIG.get("stt_compute_type")
- device, compute_type = _resolve_faster_whisper_runtime(device_value, compute_value)
+ device, compute_type = _resolve_faster_whisper_runtime("auto", "auto")
whisper_model = _load_faster_whisper_model(
model_name=model,
device=device,
compute_type=compute_type,
)
- segments, _info = whisper_model.transcribe(
- str(audio_path),
- language=language or None,
- beam_size=1,
- condition_on_previous_text=False,
- )
- transcript = re.sub(
- r"\s+",
- " ",
- " ".join(str(segment.text or "").strip() for segment in segments),
- ).strip()
+ with FASTER_WHISPER_TRANSCRIBE_LOCK:
+ segments, _info = whisper_model.transcribe(
+ str(audio_path),
+ language=language or None,
+ beam_size=1,
+ condition_on_previous_text=False,
+ )
+ transcript = re.sub(
+ r"\s+",
+ " ",
+ " ".join(str(segment.text or "").strip() for segment in segments),
+ ).strip()
with AUTO_TRAIN_LOCK:
+ AUTO_TRAIN_STATE["last_stt_engine"] = STT_ENGINE_FASTER_WHISPER
+ AUTO_TRAIN_STATE["last_stt_model"] = model
AUTO_TRAIN_STATE["last_stt_device"] = device
AUTO_TRAIN_STATE["last_stt_compute_type"] = compute_type
_save_auto_train_state_locked()
return transcript
+def _parakeet_onnx_providers() -> List[str]:
+ try:
+ import onnxruntime as ort
+ except Exception as exc:
+ raise RuntimeError(f"onnxruntime is unavailable: {exc}") from exc
+ available = [str(value) for value in ort.get_available_providers()]
+ preferred = [
+ "CUDAExecutionProvider",
+ "CPUExecutionProvider",
+ ]
+ resolved = [provider for provider in preferred if provider in set(available)]
+ if not resolved:
+ raise RuntimeError("ONNX Runtime has no usable CUDA or CPU execution provider.")
+ return resolved
+
+
+def _load_parakeet_onnx_model():
+ try:
+ import onnx_asr
+ except Exception as exc:
+ raise RuntimeError(f"onnx-asr is unavailable: {exc}") from exc
+
+ providers = tuple(_parakeet_onnx_providers())
+ cache_key = (
+ DEFAULT_PARAKEET_ONNX_MODEL,
+ DEFAULT_PARAKEET_ONNX_QUANTIZATION,
+ providers,
+ )
+ with PARAKEET_ONNX_MODEL_LOCK:
+ cached = PARAKEET_ONNX_MODEL_CACHE.get(cache_key)
+ if cached is not None:
+ return cached
+ AUTO_TRAIN_MODEL_DIR.mkdir(parents=True, exist_ok=True)
+ previous = {
+ key: os.environ.get(key)
+ for key in ("HF_HOME", "HF_HUB_CACHE", "HUGGINGFACE_HUB_CACHE")
+ }
+ os.environ["HF_HOME"] = str(AUTO_TRAIN_MODEL_DIR)
+ os.environ["HF_HUB_CACHE"] = str(AUTO_TRAIN_MODEL_DIR / "hub")
+ os.environ["HUGGINGFACE_HUB_CACHE"] = str(AUTO_TRAIN_MODEL_DIR / "hub")
+ try:
+ model = onnx_asr.load_model(
+ DEFAULT_PARAKEET_ONNX_MODEL,
+ str(AUTO_TRAIN_MODEL_DIR),
+ quantization=DEFAULT_PARAKEET_ONNX_QUANTIZATION,
+ providers=list(providers),
+ )
+ finally:
+ for key, value in previous.items():
+ if value is None:
+ os.environ.pop(key, None)
+ else:
+ os.environ[key] = value
+ PARAKEET_ONNX_MODEL_CACHE.clear()
+ PARAKEET_ONNX_MODEL_CACHE[cache_key] = model
+ return model
+
+
+def _normalized_wav_float32(audio_path: Path):
+ import numpy as np
+
+ with wave.open(str(audio_path), "rb") as wav_file:
+ channels = wav_file.getnchannels()
+ sample_width = wav_file.getsampwidth()
+ sample_rate = wav_file.getframerate()
+ frames = wav_file.readframes(wav_file.getnframes())
+ if sample_width != 2 or sample_rate != TARGET_SAMPLE_RATE or channels < 1:
+ raise RuntimeError("STT input must be 16 kHz, 16-bit PCM WAV audio.")
+ samples = np.frombuffer(frames, dtype=np.int16).astype(np.float32)
+ if channels > 1:
+ samples = samples.reshape((-1, channels)).mean(axis=1)
+ return samples / 32768.0
+
+
+def _transcribe_capture_with_parakeet(audio_path: Path, *, model: str, language: str) -> str:
+ parakeet_model = _load_parakeet_onnx_model()
+ kwargs: Dict[str, Any] = {
+ "sample_rate": TARGET_SAMPLE_RATE,
+ "channel": "mean",
+ }
+ if language:
+ kwargs["language"] = language
+ with PARAKEET_ONNX_TRANSCRIBE_LOCK:
+ result = parakeet_model.recognize(
+ _normalized_wav_float32(audio_path),
+ **kwargs,
+ )
+ providers = _parakeet_onnx_providers()
+ with AUTO_TRAIN_LOCK:
+ AUTO_TRAIN_STATE["last_stt_engine"] = STT_ENGINE_PARAKEET_ONNX
+ AUTO_TRAIN_STATE["last_stt_model"] = model
+ AUTO_TRAIN_STATE["last_stt_device"] = providers[0]
+ AUTO_TRAIN_STATE["last_stt_compute_type"] = DEFAULT_PARAKEET_ONNX_QUANTIZATION
+ _save_auto_train_state_locked()
+ return re.sub(r"\s+", " ", str(result or "")).strip()
+
+
+def _transcribe_capture(audio_path: Path, *, engine: str, language: str) -> str:
+ token = _normalize_stt_engine(engine)
+ model = _managed_stt_model(token, language)
+ if token == STT_ENGINE_PARAKEET_ONNX:
+ return _transcribe_capture_with_parakeet(
+ audio_path,
+ model=model,
+ language=language,
+ )
+ return _transcribe_capture_with_faster_whisper(
+ audio_path,
+ model=model,
+ language=language,
+ )
+
+
+def _clear_stt_model_caches(*, keep_engine: str) -> None:
+ token = _normalize_stt_engine(keep_engine)
+ cleared = False
+ if token != STT_ENGINE_FASTER_WHISPER:
+ with FASTER_WHISPER_TRANSCRIBE_LOCK:
+ with FASTER_WHISPER_MODEL_LOCK:
+ cleared = bool(FASTER_WHISPER_MODEL_CACHE) or cleared
+ FASTER_WHISPER_MODEL_CACHE.clear()
+ if token != STT_ENGINE_PARAKEET_ONNX:
+ with PARAKEET_ONNX_TRANSCRIBE_LOCK:
+ with PARAKEET_ONNX_MODEL_LOCK:
+ cleared = bool(PARAKEET_ONNX_MODEL_CACHE) or cleared
+ PARAKEET_ONNX_MODEL_CACHE.clear()
+ if cleared:
+ gc.collect()
+
+
def _queue_auto_review(file_name: str) -> bool:
safe_file_name = Path(str(file_name or "")).name
if not safe_file_name:
@@ -960,12 +1156,17 @@ def _auto_review_capture(file_name: str) -> None:
metadata["auto_review_status"] = "transcribing"
metadata["auto_reviewed_at"] = _iso_now()
metadata["auto_review_wake_phrase"] = wake_phrase
- metadata["auto_review_stt_model"] = config["stt_model"]
+ stt_engine = _normalize_stt_engine(config.get("stt_engine"))
+ metadata["auto_review_stt_engine"] = stt_engine
+ metadata["auto_review_stt_model"] = _managed_stt_model(
+ stt_engine,
+ config.get("language"),
+ )
_write_sidecar_json(audio_path, metadata)
- transcript = _transcribe_capture_with_faster_whisper(
+ transcript = _transcribe_capture(
audio_path,
- model=str(config["stt_model"]),
+ engine=stt_engine,
language=str(config.get("language") or DEFAULT_LANGUAGE),
)
normalized = _normalize_transcript_text(transcript)
@@ -2386,7 +2587,7 @@ def auto_train_status(request: Request):
payload = _auto_train_status_payload()
payload["ok"] = True
payload["advertised_base_url"] = _advertised_base_url(request)
- payload["stt_backend"] = "faster-whisper"
+ payload["stt_backend"] = payload["config"].get("stt_engine")
return payload
@@ -2420,6 +2621,8 @@ def update_auto_train(payload: Dict[str, Any] = None):
)
if schedule_changed or not AUTO_TRAIN_STATE.get("next_run_at"):
_schedule_next_auto_run_locked()
+ if previous.get("stt_engine") != normalized.get("stt_engine"):
+ _clear_stt_model_caches(keep_engine=normalized["stt_engine"])
if normalized["enabled"]:
queued = _queue_pending_auto_reviews()
AUTO_TRAIN_WAKE_EVENT.set()