""" Speech-to-Text engines for MARK XL. Whisper – offline transcription via faster-whisper (VAD-buffered) Vosk – offline streaming transcription (lighter) """ import json import numpy as np class WhisperSTT: """Offline transcription using faster-whisper.""" def __init__(self, model_name: str = "base", language: str | None = None): import os from faster_whisper import WhisperModel print(f"[STT] Loading Whisper '{model_name}'…") try: import torch device = "cuda" if torch.cuda.is_available() else "cpu" compute = "float16" if device == "cuda" else "int8" except Exception: device, compute = "cpu", "int8" try: self._model = WhisperModel(model_name, device=device, compute_type=compute) except Exception as _first_err: # Offline flag set but model not cached yet → clear flags and download once. # Keywords cover multiple huggingface_hub error message variants across versions. _e = str(_first_err).lower() _offline_keywords = ( "offline", "not found", "cache", "localentry", "does not exist", "outgoing", "local_files_only", ) if any(k in _e for k in _offline_keywords): print(f"[STT] Whisper '{model_name}' not in local cache — downloading (one-time, internet required)…") os.environ.pop("HF_HUB_OFFLINE", None) os.environ.pop("TRANSFORMERS_OFFLINE", None) os.environ.pop("HF_DATASETS_OFFLINE", None) try: self._model = WhisperModel(model_name, device=device, compute_type=compute) except Exception as _dl_err: raise RuntimeError( f"Whisper '{model_name}' model download failed.\n" f"Internet access is required the first time to download the speech model (~75–290 MB).\n" f"After the first download it runs fully offline.\n" f"Details: {_dl_err}" ) from _dl_err else: raise self._language = None if (not language or language.strip().lower() == "auto") else language.strip().lower() print(f"[STT] Whisper '{model_name}' ready ({device})") def transcribe(self, audio: np.ndarray) -> str: """Transcribe a float32 mono 16 kHz numpy array. Returns transcript string.""" try: segments, _ = self._model.transcribe( audio, language=self._language, beam_size=1, # greedy — 2-3x faster best_of=1, condition_on_previous_text=False, # no hallucinations, faster vad_filter=True, vad_parameters={"min_silence_duration_ms": 300}, ) return " ".join(s.text for s in segments).strip() except Exception as e: print(f"[STT] Transcription error: {e}") raise class VoskSTT: """Streaming transcription using Vosk.""" def __init__(self, model_path: str | None = None, language: str = "en-us"): from vosk import Model, KaldiRecognizer print("[STT] Loading Vosk model…") if model_path: model = Model(model_path) else: lang = language.strip().lower() if language and language.strip().lower() != "auto" else "en-us" model = Model(lang=lang) self._rec = KaldiRecognizer(model, 16000) print("[STT] Vosk ready.") def process_chunk(self, audio_bytes: bytes) -> tuple[str, bool]: """Feed raw int16 LE PCM bytes. Returns (text, is_final).""" if self._rec.AcceptWaveform(audio_bytes): result = json.loads(self._rec.Result()) return result.get("text", ""), True partial = json.loads(self._rec.PartialResult()) return partial.get("partial", ""), False