Files
whisper_voice/src/core/transcriber.py
2026-01-25 20:28:01 +02:00

253 lines
10 KiB
Python

"""
Whisper Transcriber Module.
===========================
Transcriber Module.
===================
Handles audio transcription using faster-whisper.
Runs IN-PROCESS (no subprocess) to ensure stability on all systems.
"""
import os
import logging
from typing import Optional
import numpy as np
from src.core.config import ConfigManager
from src.core.paths import get_models_path
try:
import torch
except ImportError:
torch = None
# Import directly - valid since we are now running in the full environment
class WhisperTranscriber:
"""
Manages the faster-whisper model and transcription process.
"""
def __init__(self):
"""Initialize settings."""
self.config = ConfigManager()
self.model = None
self.current_model_size = None
self.current_compute_device = None
self.current_compute_type = None
def load_model(self):
"""
Loads the model specified in config.
Safe to call multiple times (checks if reload needed).
"""
size = self.config.get("model_size")
device = self.config.get("compute_device")
compute = self.config.get("compute_type")
# Check if already loaded
if (self.model and
self.current_model_size == size and
self.current_compute_device == device and
self.current_compute_type == compute):
return
logging.info(f"Loading Model: {size} on {device} ({compute})...")
try:
# Construct path to local model for offline support
new_path = get_models_path() / f"faster-whisper-{size}"
model_input = str(new_path) if new_path.exists() else size
# Force offline if path exists to avoid HF errors
local_only = new_path.exists()
try:
from faster_whisper import WhisperModel
self.model = WhisperModel(
model_input,
device=device,
compute_type=compute,
download_root=str(get_models_path()),
local_files_only=local_only
)
except Exception as load_err:
# CRITICAL FALLBACK: If CUDA/cublas fails (AMD/Intel users), fallback to CPU
err_str = str(load_err).lower()
if "cublas" in err_str or "cudnn" in err_str or "library" in err_str or "device" in err_str:
logging.warning(f"CUDA Init Failed ({load_err}). Falling back to CPU...")
self.config.set("compute_device", "cpu") # Update config for persistence/UI
self.current_compute_device = "cpu"
self.model = WhisperModel(
model_input,
device="cpu",
compute_type="int8", # CPU usually handles int8 well with newer extensions, or standard
download_root=str(get_models_path()),
local_files_only=local_only
)
else:
raise load_err
self.current_model_size = size
self.current_compute_device = device
self.current_compute_type = compute
logging.info("Model loaded successfully.")
except Exception as e:
logging.error(f"Failed to load model: {e}")
self.model = None
# Auto-Repair: Detect vocabulary/corrupt errors
err_str = str(e).lower()
if "vocabulary" in err_str or "tokenizer" in err_str or "config.json" in err_str:
# ... existing auto-repair logic ...
logging.warning("Corrupt model detected on load. Attempting to delete and reset...")
try:
import shutil
# Differentiate between simple path and HF path
new_path = get_models_path() / f"faster-whisper-{size}"
if new_path.exists():
shutil.rmtree(new_path)
logging.info(f"Deleted corrupt model at {new_path}")
else:
# Try legacy HF path
hf_path = get_models_path() / f"models--Systran--faster-whisper-{size}"
if hf_path.exists():
shutil.rmtree(hf_path)
logging.info(f"Deleted corrupt HF model at {hf_path}")
# Notify UI to refresh state (will show 'Download' button now)
# We can't reach bridge easily here without passing it in,
# but the UI polls or listens to logs.
# The user will simply see "Model Missing" in settings after this.
except Exception as del_err:
logging.error(f"Failed to delete corrupt model: {del_err}")
def transcribe(self, audio_data, is_file: bool = False, task: Optional[str] = None) -> str:
"""
Transcribe audio data.
"""
logging.info(f"Starting transcription... (is_file={is_file}, task={task})")
# Ensure model is loaded
if not self.model:
self.load_model()
if not self.model:
return "Error: Model failed to load. Please check Settings -> Model Info."
try:
# Config
beam_size = int(self.config.get("beam_size"))
best_of = int(self.config.get("best_of"))
vad = False if is_file else self.config.get("vad_filter")
language = self.config.get("language")
# Use task override if provided, otherwise config
# Ensure safe string and lowercase ("transcribe" vs "Transcribe")
raw_task = task if task else self.config.get("task")
final_task = str(raw_task).strip().lower() if raw_task else "transcribe"
# Sanity check for valid Whisper tasks
if final_task not in ["transcribe", "translate"]:
logging.warning(f"Invalid task '{final_task}' detected. Defaulting to 'transcribe'.")
final_task = "transcribe"
# Language handling
final_language = language if language != "auto" else None
# Anti-Hallucination: Force condition_on_previous_text=False for translation
condition_prev = self.config.get("condition_on_previous_text")
# Helper options for Translation Stability
initial_prompt = self.config.get("initial_prompt")
if final_task == "translate":
condition_prev = False
# Force beam search if user has set it to greedy (1)
# Translation requires more search breadth to find the English mapping
if beam_size < 5:
logging.info("Forcing beam_size=5 for Translation task.")
beam_size = 5
# Inject guidance prompt if none exists
if not initial_prompt:
initial_prompt = "Translate this to English."
logging.info(f"Model Dispatch: Task='{final_task}', Language='{final_language}', ConditionPrev={condition_prev}, Beam={beam_size}")
# Build arguments dynamically to avoid passing None if that's the issue
transcribe_opts = {
"beam_size": beam_size,
"best_of": best_of,
"vad_filter": vad,
"task": final_task,
"vad_parameters": dict(min_silence_duration_ms=500),
"condition_on_previous_text": condition_prev,
"without_timestamps": True
}
if initial_prompt:
transcribe_opts["initial_prompt"] = initial_prompt
# Only add language if it's explicitly set (not None/Auto)
# This avoids potentially confusing the model with explicit None
if final_language:
transcribe_opts["language"] = final_language
# Transcribe
segments, info = self.model.transcribe(audio_data, **transcribe_opts)
# Aggregate text
text_result = ""
for segment in segments:
text_result += segment.text + " "
text_result = text_result.strip()
# Low VRAM Mode: Unload Whisper Model immediately
if self.config.get("unload_models_after_use"):
self.unload_model()
logging.info(f"Final Transcription Output: '{text_result}'")
return text_result
except Exception as e:
logging.error(f"Transcription failed: {e}")
return f"Error: {str(e)}"
def model_exists(self, size: str) -> bool:
"""Checks if a model size is already downloaded."""
new_path = get_models_path() / f"faster-whisper-{size}"
if new_path.exists():
# Strict check
required = ["config.json", "model.bin", "vocabulary.json"]
if all((new_path / f).exists() for f in required):
return True
# Legacy HF cache check
folder_name = f"models--Systran--faster-whisper-{size}"
path = get_models_path() / folder_name / "snapshots"
if path.exists() and any(path.iterdir()):
return True
return False
def unload_model(self):
"""
Unloads model to free memory.
"""
if self.model:
del self.model
self.model = None
self.current_model_size = None
# Force garbage collection
import gc
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
logging.info("Whisper Model unloaded (Low VRAM Mode).")