TalklasApp / app.py
Jerich's picture
Fix TTS language issue by dynamically updating model in /translate-audio endpoint
2b35bda verified
import os
os.environ["HOME"] = "/root"
os.environ["HF_HOME"] = "/tmp/hf_cache"
import logging
import threading
import tempfile
import uuid
import torch
import numpy as np
import soundfile as sf
import torchaudio
import wave
import time
from fastapi import FastAPI, HTTPException, UploadFile, File, Form, BackgroundTasks
from fastapi.responses import JSONResponse
from fastapi.staticfiles import StaticFiles
from typing import Dict, Any, Optional, Tuple
from datetime import datetime, timedelta
# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger("talklas-api")
app = FastAPI(title="Talklas API")
# Mount a directory to serve audio files
AUDIO_DIR = "/tmp/audio_output" # Use /tmp for temporary files
os.makedirs(AUDIO_DIR, exist_ok=True)
app.mount("/audio_output", StaticFiles(directory=AUDIO_DIR), name="audio_output")
# Global variables to track application state
models_loaded = False
loading_in_progress = False
loading_thread = None
model_status = {
"stt": "not_loaded",
"mt": "not_loaded",
"tts": "not_loaded"
}
error_message = None
current_tts_language = "tgl" # Track the current TTS language
# Model instances
stt_processor = None
stt_model = None
mt_model = None
mt_tokenizer = None
tts_model = None
tts_tokenizer = None
# Define the valid languages and mappings
LANGUAGE_MAPPING = {
"English": "eng",
"Tagalog": "tgl",
"Cebuano": "ceb",
"Ilocano": "ilo",
"Waray": "war",
"Pangasinan": "pag"
}
NLLB_LANGUAGE_CODES = {
"eng": "eng_Latn",
"tgl": "tgl_Latn",
"ceb": "ceb_Latn",
"ilo": "ilo_Latn",
"war": "war_Latn",
"pag": "pag_Latn"
}
# Function to save PCM data as a WAV file
def save_pcm_to_wav(pcm_data: list, sample_rate: int, output_path: str):
# Convert pcm_data to a NumPy array of 16-bit integers
pcm_array = np.array(pcm_data, dtype=np.int16)
with wave.open(output_path, 'wb') as wav_file:
# Set WAV parameters: 1 channel (mono), 2 bytes per sample (16-bit), sample rate
wav_file.setnchannels(1)
wav_file.setsampwidth(2) # 16-bit audio
wav_file.setframerate(sample_rate)
# Write the 16-bit PCM data as bytes (little-endian)
wav_file.writeframes(pcm_array.tobytes())
# Function to detect speech using an energy-based approach
def detect_speech(waveform: torch.Tensor, sample_rate: int, threshold: float = 0.01, min_speech_duration: float = 0.5) -> bool:
"""
Detects if the audio contains speech using an energy-based approach.
Returns True if speech is detected, False otherwise.
"""
# Convert waveform to numpy array
waveform_np = waveform.numpy()
if waveform_np.ndim > 1:
waveform_np = waveform_np.mean(axis=0) # Convert stereo to mono
# Compute RMS energy
rms = np.sqrt(np.mean(waveform_np**2))
logger.info(f"RMS energy: {rms}")
# Check if RMS energy exceeds the threshold
if rms < threshold:
logger.info("No speech detected: RMS energy below threshold")
return False
# Optionally, check for minimum speech duration (requires more sophisticated VAD)
# For now, we assume if RMS is above threshold, there is speech
return True
# Function to clean up old audio files
def cleanup_old_audio_files():
logger.info("Starting cleanup of old audio files...")
expiration_time = datetime.now() - timedelta(minutes=10) # Files older than 10 minutes
for filename in os.listdir(AUDIO_DIR):
file_path = os.path.join(AUDIO_DIR, filename)
if os.path.isfile(file_path):
file_mtime = datetime.fromtimestamp(os.path.getmtime(file_path))
if file_mtime < expiration_time:
try:
os.unlink(file_path)
logger.info(f"Deleted old audio file: {file_path}")
except Exception as e:
logger.error(f"Error deleting file {file_path}: {str(e)}")
# Background task to periodically clean up audio files
def schedule_cleanup():
while True:
cleanup_old_audio_files()
time.sleep(300) # Run every 5 minutes (300 seconds)
# Function to load models in background
def load_models_task():
global models_loaded, loading_in_progress, model_status, error_message
global stt_processor, stt_model, mt_model, mt_tokenizer, tts_model, tts_tokenizer
try:
loading_in_progress = True
# Load STT model (MMS with fallback to Whisper)
logger.info("Starting to load STT model...")
from transformers import AutoProcessor, AutoModelForCTC, WhisperProcessor, WhisperForConditionalGeneration
try:
logger.info("Loading MMS STT model...")
model_status["stt"] = "loading"
stt_processor = AutoProcessor.from_pretrained("facebook/mms-1b-all")
stt_model = AutoModelForCTC.from_pretrained("facebook/mms-1b-all")
device = "cuda" if torch.cuda.is_available() else "cpu"
stt_model.to(device)
logger.info("MMS STT model loaded successfully")
model_status["stt"] = "loaded_mms"
except Exception as mms_error:
logger.error(f"Failed to load MMS STT model: {str(mms_error)}")
logger.info("Falling back to Whisper STT model...")
try:
stt_processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
stt_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")
stt_model.to(device)
logger.info("Whisper STT model loaded successfully as fallback")
model_status["stt"] = "loaded_whisper"
except Exception as whisper_error:
logger.error(f"Failed to load Whisper STT model: {str(whisper_error)}")
model_status["stt"] = "failed"
error_message = f"STT model loading failed: MMS error: {str(mms_error)}, Whisper error: {str(whisper_error)}"
return
# Load MT model
logger.info("Starting to load MT model...")
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
try:
logger.info("Loading NLLB-200-distilled-600M model...")
model_status["mt"] = "loading"
mt_model = AutoModelForSeq2SeqLM.from_pretrained("facebook/nllb-200-distilled-600M")
mt_tokenizer = AutoTokenizer.from_pretrained("facebook/nllb-200-distilled-600M")
mt_model.to(device)
logger.info("MT model loaded successfully")
model_status["mt"] = "loaded"
except Exception as e:
logger.error(f"Failed to load MT model: {str(e)}")
model_status["mt"] = "failed"
error_message = f"MT model loading failed: {str(e)}"
return
# Load TTS model (default to Tagalog, will be updated dynamically)
logger.info("Starting to load TTS model...")
from transformers import VitsModel, AutoTokenizer
try:
logger.info("Loading MMS-TTS model for Tagalog...")
model_status["tts"] = "loading"
tts_model = VitsModel.from_pretrained("facebook/mms-tts-tgl")
tts_tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-tgl")
tts_model.to(device)
logger.info("TTS model loaded successfully")
model_status["tts"] = "loaded"
except Exception as e:
logger.error(f"Failed to load TTS model for Tagalog: {str(e)}")
# Fallback to English TTS if the target language fails
try:
logger.info("Falling back to MMS-TTS English model...")
tts_model = VitsModel.from_pretrained("facebook/mms-tts-eng")
tts_tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-eng")
tts_model.to(device)
logger.info("Fallback TTS model loaded successfully")
model_status["tts"] = "loaded (fallback)"
current_tts_language = "eng"
except Exception as e2:
logger.error(f"Failed to load fallback TTS model: {str(e2)}")
model_status["tts"] = "failed"
error_message = f"TTS model loading failed: {str(e)} (fallback also failed: {str(e2)})"
return
models_loaded = True
logger.info("Model loading completed successfully")
except Exception as e:
error_message = str(e)
logger.error(f"Error in model loading task: {str(e)}")
finally:
loading_in_progress = False
# Start loading models in background
def start_model_loading():
global loading_thread, loading_in_progress
if not loading_in_progress and not models_loaded:
loading_in_progress = True
loading_thread = threading.Thread(target=load_models_task)
loading_thread.daemon = True
loading_thread.start()
# Start the background cleanup task
def start_cleanup_task():
cleanup_thread = threading.Thread(target=schedule_cleanup)
cleanup_thread.daemon = True
cleanup_thread.start()
# Start the background processes when the app starts
@app.on_event("startup")
async def startup_event():
logger.info("Application starting up...")
start_model_loading()
start_cleanup_task()
@app.get("/")
async def root():
"""Root endpoint for default health check"""
logger.info("Root endpoint requested")
return {"status": "healthy"}
@app.get("/health")
async def health_check():
"""Health check endpoint that always returns successfully"""
global models_loaded, loading_in_progress, model_status, error_message
logger.info("Health check requested")
return {
"status": "healthy",
"models_loaded": models_loaded,
"loading_in_progress": loading_in_progress,
"model_status": model_status,
"error": error_message
}
@app.post("/update-languages")
async def update_languages(source_lang: str = Form(...), target_lang: str = Form(...)):
global stt_processor, stt_model, tts_model, tts_tokenizer, current_tts_language
if source_lang not in LANGUAGE_MAPPING or target_lang not in LANGUAGE_MAPPING:
raise HTTPException(status_code=400, detail="Invalid language selected")
source_code = LANGUAGE_MAPPING[source_lang]
target_code = LANGUAGE_MAPPING[target_lang]
# Update the STT model based on the source language (MMS or Whisper)
try:
logger.info("Updating STT model for source language...")
from transformers import AutoProcessor, AutoModelForCTC, WhisperProcessor, WhisperForConditionalGeneration
device = "cuda" if torch.cuda.is_available() else "cpu"
try:
logger.info(f"Loading MMS STT model for {source_code}...")
stt_processor = AutoProcessor.from_pretrained("facebook/mms-1b-all")
stt_model = AutoModelForCTC.from_pretrained("facebook/mms-1b-all")
stt_model.to(device)
# Set the target language for MMS
if source_code in stt_processor.tokenizer.vocab.keys():
stt_processor.tokenizer.set_target_lang(source_code)
stt_model.load_adapter(source_code)
logger.info(f"MMS STT model updated to {source_code}")
model_status["stt"] = "loaded_mms"
else:
logger.warning(f"Language {source_code} not supported by MMS, using default")
model_status["stt"] = "loaded_mms_default"
except Exception as mms_error:
logger.error(f"Failed to load MMS STT model for {source_code}: {str(mms_error)}")
logger.info("Falling back to Whisper STT model...")
try:
stt_processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
stt_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")
stt_model.to(device)
logger.info("Whisper STT model loaded successfully as fallback")
model_status["stt"] = "loaded_whisper"
except Exception as whisper_error:
logger.error(f"Failed to load Whisper STT model: {str(whisper_error)}")
model_status["stt"] = "failed"
error_message = f"STT model update failed: MMS error: {str(mms_error)}, Whisper error: {str(whisper_error)}"
return {"status": "failed", "error": error_message}
except Exception as e:
logger.error(f"Error updating STT model: {str(e)}")
model_status["stt"] = "failed"
error_message = f"STT model update failed: {str(e)}"
return {"status": "failed", "error": error_message}
# Update the TTS model based on the target language
try:
logger.info(f"Loading MMS-TTS model for {target_code}...")
from transformers import VitsModel, AutoTokenizer
tts_model = VitsModel.from_pretrained(f"facebook/mms-tts-{target_code}")
tts_tokenizer = AutoTokenizer.from_pretrained(f"facebook/mms-tts-{target_code}")
tts_model.to(device)
current_tts_language = target_code
logger.info(f"TTS model updated to {target_code}")
model_status["tts"] = "loaded"
except Exception as e:
logger.error(f"Failed to load TTS model for {target_code}: {str(e)}")
try:
logger.info("Falling back to MMS-TTS English model...")
tts_model = VitsModel.from_pretrained("facebook/mms-tts-eng")
tts_tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-eng")
tts_model.to(device)
current_tts_language = "eng"
logger.info("Fallback TTS model loaded successfully")
model_status["tts"] = "loaded (fallback)"
except Exception as e2:
logger.error(f"Failed to load fallback TTS model: {str(e2)}")
model_status["tts"] = "failed"
error_message = f"TTS model loading failed: {str(e)} (fallback also failed: {str(e2)})"
return {"status": "failed", "error": error_message}
logger.info(f"Updating languages: {source_lang}{target_lang}")
return {"status": f"Languages updated to {source_lang}{target_lang}"}
@app.post("/translate-text")
async def translate_text(text: str = Form(...), source_lang: str = Form(...), target_lang: str = Form(...)):
"""Endpoint to translate text and convert to speech"""
global mt_model, mt_tokenizer, tts_model, tts_tokenizer, current_tts_language
if not text:
raise HTTPException(status_code=400, detail="No text provided")
if source_lang not in LANGUAGE_MAPPING or target_lang not in LANGUAGE_MAPPING:
raise HTTPException(status_code=400, detail="Invalid language selected")
logger.info(f"Translate-text requested: {text} from {source_lang} to {target_lang}")
request_id = str(uuid.uuid4())
# Translate the text
source_code = LANGUAGE_MAPPING[source_lang]
target_code = LANGUAGE_MAPPING[target_lang]
translated_text = "Translation not available"
if model_status["mt"] == "loaded" and mt_model is not None and mt_tokenizer is not None:
try:
source_nllb_code = NLLB_LANGUAGE_CODES[source_code]
target_nllb_code = NLLB_LANGUAGE_CODES[target_code]
mt_tokenizer.src_lang = source_nllb_code
device = "cuda" if torch.cuda.is_available() else "cpu"
inputs = mt_tokenizer(text, return_tensors="pt").to(device)
with torch.no_grad():
generated_tokens = mt_model.generate(
**inputs,
forced_bos_token_id=mt_tokenizer.convert_tokens_to_ids(target_nllb_code),
max_length=448
)
translated_text = mt_tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]
logger.info(f"Translation completed: {translated_text}")
except Exception as e:
logger.error(f"Error during translation: {str(e)}")
translated_text = f"Translation failed: {str(e)}"
else:
logger.warning("MT model not loaded, skipping translation")
# Update TTS model if the target language doesn't match the current TTS language
if current_tts_language != target_code:
try:
logger.info(f"Updating TTS model for {target_code}...")
from transformers import VitsModel, AutoTokenizer
tts_model = VitsModel.from_pretrained(f"facebook/mms-tts-{target_code}")
tts_tokenizer = AutoTokenizer.from_pretrained(f"facebook/mms-tts-{target_code}")
tts_model.to(device)
current_tts_language = target_code
logger.info(f"TTS model updated to {target_code}")
model_status["tts"] = "loaded"
except Exception as e:
logger.error(f"Failed to load TTS model for {target_code}: {str(e)}")
try:
logger.info("Falling back to MMS-TTS English model...")
tts_model = VitsModel.from_pretrained("facebook/mms-tts-eng")
tts_tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-eng")
tts_model.to(device)
current_tts_language = "eng"
logger.info("Fallback TTS model loaded successfully")
model_status["tts"] = "loaded (fallback)"
except Exception as e2:
logger.error(f"Failed to load fallback TTS model: {str(e2)}")
model_status["tts"] = "failed"
# Convert translated text to speech
output_audio_url = None
if model_status["tts"].startswith("loaded") and tts_model is not None and tts_tokenizer is not None:
try:
inputs = tts_tokenizer(translated_text, return_tensors="pt").to(device)
with torch.no_grad():
output = tts_model(**inputs)
speech = output.waveform.cpu().numpy().squeeze()
speech = (speech * 32767).astype(np.int16)
sample_rate = tts_model.config.sampling_rate
# Save the audio as a WAV file
output_filename = f"{request_id}.wav"
output_path = os.path.join(AUDIO_DIR, output_filename)
save_pcm_to_wav(speech.tolist(), sample_rate, output_path)
logger.info(f"Saved synthesized audio to {output_path}")
# Generate a URL to the WAV file
output_audio_url = f"https://jerich-talklasapp.hf.space/audio_output/{output_filename}"
logger.info("TTS conversion completed")
except Exception as e:
logger.error(f"Error during TTS conversion: {str(e)}")
output_audio_url = None
return {
"request_id": request_id,
"status": "completed",
"message": "Translation and TTS completed (or partially completed).",
"source_text": text,
"translated_text": translated_text,
"output_audio": output_audio_url
}
@app.post("/translate-audio")
async def translate_audio(audio: UploadFile = File(...), source_lang: str = Form(...), target_lang: str = Form(...)):
"""Endpoint to transcribe, translate, and convert audio to speech"""
global stt_processor, stt_model, mt_model, mt_tokenizer, tts_model, tts_tokenizer, current_tts_language
if not audio:
raise HTTPException(status_code=400, detail="No audio file provided")
if source_lang not in LANGUAGE_MAPPING or target_lang not in LANGUAGE_MAPPING:
raise HTTPException(status_code=400, detail="Invalid language selected")
logger.info(f"Translate-audio requested: {audio.filename} from {source_lang} to {target_lang}")
request_id = str(uuid.uuid4())
# Check if STT model is loaded
if model_status["stt"] not in ["loaded_mms", "loaded_mms_default", "loaded_whisper"] or stt_processor is None or stt_model is None:
logger.warning("STT model not loaded, returning placeholder response")
return {
"request_id": request_id,
"status": "processing",
"message": "STT model not loaded yet. Please try again later.",
"source_text": "Transcription not available",
"translated_text": "Translation not available",
"output_audio": None
}
# Save the uploaded audio to a temporary file
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_file:
temp_file.write(await audio.read())
temp_path = temp_file.name
transcription = "Transcription not available"
translated_text = "Translation not available"
output_audio_url = None
try:
# Step 1: Load and resample the audio using torchaudio
logger.info(f"Reading audio file: {temp_path}")
waveform, sample_rate = torchaudio.load(temp_path)
logger.info(f"Audio loaded: sample_rate={sample_rate}, waveform_shape={waveform.shape}")
# Resample to 16 kHz if needed (required by Whisper and MMS models)
if sample_rate != 16000:
logger.info(f"Resampling audio from {sample_rate} Hz to 16000 Hz")
resampler = torchaudio.transforms.Resample(sample_rate, 16000)
waveform = resampler(waveform)
sample_rate = 16000
# Step 2: Detect speech
if not detect_speech(waveform, sample_rate):
return {
"request_id": request_id,
"status": "failed",
"message": "No speech detected in the audio.",
"source_text": "No speech detected",
"translated_text": "No translation available",
"output_audio": None
}
# Step 3: Transcribe the audio (STT)
device = "cuda" if torch.cuda.is_available() else "cpu"
logger.info(f"Using device: {device}")
inputs = stt_processor(waveform.numpy(), sampling_rate=16000, return_tensors="pt").to(device)
logger.info("Audio processed, generating transcription...")
with torch.no_grad():
if model_status["stt"] == "loaded_whisper":
# Whisper model
generated_ids = stt_model.generate(**inputs, language="en")
transcription = stt_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
else:
# MMS model
logits = stt_model(**inputs).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = stt_processor.batch_decode(predicted_ids)[0]
logger.info(f"Transcription completed: {transcription}")
# Step 4: Translate the transcribed text (MT)
source_code = LANGUAGE_MAPPING[source_lang]
target_code = LANGUAGE_MAPPING[target_lang]
if model_status["mt"] == "loaded" and mt_model is not None and mt_tokenizer is not None:
try:
source_nllb_code = NLLB_LANGUAGE_CODES[source_code]
target_nllb_code = NLLB_LANGUAGE_CODES[target_code]
mt_tokenizer.src_lang = source_nllb_code
inputs = mt_tokenizer(transcription, return_tensors="pt").to(device)
with torch.no_grad():
generated_tokens = mt_model.generate(
**inputs,
forced_bos_token_id=mt_tokenizer.convert_tokens_to_ids(target_nllb_code),
max_length=448
)
translated_text = mt_tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]
logger.info(f"Translation completed: {translated_text}")
except Exception as e:
logger.error(f"Error during translation: {str(e)}")
translated_text = f"Translation failed: {str(e)}"
else:
logger.warning("MT model not loaded, skipping translation")
# Step 5: Update TTS model if the target language doesn't match the current TTS language
if current_tts_language != target_code:
try:
logger.info(f"Updating TTS model for {target_code}...")
from transformers import VitsModel, AutoTokenizer
tts_model = VitsModel.from_pretrained(f"facebook/mms-tts-{target_code}")
tts_tokenizer = AutoTokenizer.from_pretrained(f"facebook/mms-tts-{target_code}")
tts_model.to(device)
current_tts_language = target_code
logger.info(f"TTS model updated to {target_code}")
model_status["tts"] = "loaded"
except Exception as e:
logger.error(f"Failed to load TTS model for {target_code}: {str(e)}")
try:
logger.info("Falling back to MMS-TTS English model...")
tts_model = VitsModel.from_pretrained("facebook/mms-tts-eng")
tts_tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-eng")
tts_model.to(device)
current_tts_language = "eng"
logger.info("Fallback TTS model loaded successfully")
model_status["tts"] = "loaded (fallback)"
except Exception as e2:
logger.error(f"Failed to load fallback TTS model: {str(e2)}")
model_status["tts"] = "failed"
# Step 6: Convert translated text to speech (TTS)
if model_status["tts"].startswith("loaded") and tts_model is not None and tts_tokenizer is not None:
try:
inputs = tts_tokenizer(translated_text, return_tensors="pt").to(device)
with torch.no_grad():
output = tts_model(**inputs)
speech = output.waveform.cpu().numpy().squeeze()
speech = (speech * 32767).astype(np.int16)
sample_rate = tts_model.config.sampling_rate
# Save the audio as a WAV file
output_filename = f"{request_id}.wav"
output_path = os.path.join(AUDIO_DIR, output_filename)
save_pcm_to_wav(speech.tolist(), sample_rate, output_path)
logger.info(f"Saved synthesized audio to {output_path}")
# Generate a URL to the WAV file
output_audio_url = f"https://jerich-talklasapp.hf.space/audio_output/{output_filename}"
logger.info("TTS conversion completed")
except Exception as e:
logger.error(f"Error during TTS conversion: {str(e)}")
output_audio_url = None
return {
"request_id": request_id,
"status": "completed",
"message": "Transcription, translation, and TTS completed (or partially completed).",
"source_text": transcription,
"translated_text": translated_text,
"output_audio": output_audio_url
}
except Exception as e:
logger.error(f"Error during processing: {str(e)}")
return {
"request_id": request_id,
"status": "failed",
"message": f"Processing failed: {str(e)}",
"source_text": transcription,
"translated_text": translated_text,
"output_audio": output_audio_url
}
finally:
logger.info(f"Cleaning up temporary file: {temp_path}")
os.unlink(temp_path)
if __name__ == "__main__":
import uvicorn
logger.info("Starting Uvicorn server...")
uvicorn.run(app, host="0.0.0.0", port=8000, workers=1)