@staff0rd/assist 0.751.0 → 0.752.0

package/dist/commands/voice/python/smart_turn.py

@@ -1,63 +0,0 @@
-"""Smart Turn end-of-utterance detection (ONNX) via pipecat-ai/smart-turn."""
-
-import os
-
-import numpy as np
-import onnxruntime as ort
-from transformers import WhisperFeatureExtractor
-
-from logger import log
-
-END_THRESHOLD = 0.5
-CHUNK_SECONDS = 8
-SAMPLE_RATE = 16000
-
-
-def _truncate_or_pad(audio: np.ndarray) -> np.ndarray:
-    max_samples = CHUNK_SECONDS * SAMPLE_RATE
-    if len(audio) > max_samples:
-        return audio[-max_samples:]
-    if len(audio) < max_samples:
-        padding = max_samples - len(audio)
-        return np.pad(audio, (padding, 0), mode="constant", constant_values=0)
-    return audio
-
-
-class SmartTurn:
-    def __init__(self):
-        model_path = os.environ.get("VOICE_MODEL_SMART_TURN")
-        if not model_path:
-            models_dir = os.environ.get(
-                "VOICE_MODELS_DIR",
-                os.path.expanduser("~/.assist/voice/models"),
-            )
-            model_path = os.path.join(models_dir, "smart-turn-v3.2-cpu.onnx")
-
-        log("smart_turn_init", f"model={model_path}")
-        so = ort.SessionOptions()
-        so.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
-        so.inter_op_num_threads = 1
-        so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
-        self._session = ort.InferenceSession(
-            model_path, sess_options=so, providers=["CPUExecutionProvider"]
-        )
-        self._feature_extractor = WhisperFeatureExtractor(chunk_length=CHUNK_SECONDS)
-        self.threshold = END_THRESHOLD
-
-    def is_end_of_turn(self, audio: np.ndarray) -> bool:
-        """Check if the accumulated audio indicates end of utterance."""
-        audio = _truncate_or_pad(audio)
-        inputs = self._feature_extractor(
-            audio,
-            sampling_rate=SAMPLE_RATE,
-            return_tensors="np",
-            padding="max_length",
-            max_length=CHUNK_SECONDS * SAMPLE_RATE,
-            truncation=True,
-            do_normalize=True,
-        )
-        features = inputs.input_features.squeeze(0).astype(np.float32)
-        features = np.expand_dims(features, axis=0)
-        outputs = self._session.run(None, {"input_features": features})
-        prob = float(outputs[0][0].item())
-        return prob > self.threshold

package/dist/commands/voice/python/stt.py

@@ -1,54 +0,0 @@
-"""Parakeet NeMo STT wrapper (GPU)."""
-
-import os
-import numpy as np
-import torch
-
-from logger import log
-
-DEFAULT_MODEL = "nvidia/parakeet-ctc-1.1b"
-
-
-class ParakeetSTT:
-    def __init__(self):
-        model_name = os.environ.get("VOICE_MODEL_STT", DEFAULT_MODEL)
-        self._device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-        log("stt_init", f"model={model_name} device={self._device}")
-
-        import nemo.collections.asr as nemo_asr
-
-        self._model = nemo_asr.models.EncDecCTCModelBPE.from_pretrained(model_name)
-        self._model = self._model.to(self._device)
-        self._model.eval()
-        log("stt_ready")
-
-    def transcribe(self, audio: np.ndarray, sample_rate: int = 16000) -> str:
-        """Transcribe audio buffer to text via direct forward pass."""
-        audio_tensor = (
-            torch.tensor(audio, dtype=torch.float32).unsqueeze(0).to(self._device)
-        )
-        audio_len = torch.tensor([audio.shape[0]], dtype=torch.long).to(self._device)
-
-        with torch.no_grad():
-            logits, logits_len, _ = self._model.forward(
-                input_signal=audio_tensor, input_signal_length=audio_len
-            )
-            # Greedy CTC decode
-            preds = torch.argmax(logits, dim=-1)
-            text = self._model.decoding.ctc_decoder_predictions_tensor(
-                preds, decoder_lengths=logits_len
-            )
-
-        # Result may be nested: tuple of lists of Hypothesis objects
-        if isinstance(text, tuple):
-            text = text[0]
-        if isinstance(text, list):
-            text = text[0]
-        # NeMo returns Hypothesis namedtuples with a .text field
-        if hasattr(text, "text"):
-            text = text.text
-        if not isinstance(text, str):
-            text = str(text)
-
-        log("stt_result", text)
-        return text