Upload 2 files

Dockerfile

@@ -0,0 +1,34 @@
+FROM python:3.10-slim
+
+WORKDIR /app
+
+# Install system dependencies required for building and audio processing
+RUN apt-get update && apt-get install -y \
+    git bash wget build-essential libsndfile1 \
+    && rm -rf /var/lib/apt/lists/*
+
+# Force install PyTorch CPU version to save massive amounts of build time and space
+RUN pip install --no-cache-dir torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+
+# Clone and install the custom NeMo fork required by IndicConformerASR
+RUN git clone https://github.com/AI4Bharat/NeMo.git \
+    && cd NeMo \
+    && git checkout nemo-v2 \
+    && bash reinstall.sh \
+    && cd ..
+
+# Install FastAPI, Piper TTS, and patch the known dependency issues
+RUN pip install --no-cache-dir "numpy<2.0" huggingface_hub==0.23.2 fastapi uvicorn python-multipart "piper-tts==1.2.0"
+
+# Fetch the Piper Nepali Chitwan model directly during the Docker build
+RUN wget -q -O chitwan.onnx "https://huggingface.co/rhasspy/piper-voices/resolve/main/ne/ne_NP/chitwan/medium/ne_NP-chitwan-medium.onnx?download=true"
+RUN wget -q -O chitwan.onnx.json "https://huggingface.co/rhasspy/piper-voices/resolve/main/ne/ne_NP/chitwan/medium/ne_NP-chitwan-medium.onnx.json?download=true"
+
+# Copy the FastAPI app into the container
+COPY app.py /app/app.py
+
+# Hugging Face Spaces route traffic through port 7860
+EXPOSE 7860
+
+# Start the asynchronous Uvicorn server
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]

app.py

@@ -0,0 +1,100 @@
+import os
+import wave
+import asyncio
+import torch
+from fastapi import FastAPI, Depends, HTTPException, status, UploadFile, File
+from fastapi.security import HTTPBasic, HTTPBasicCredentials
+from fastapi.responses import FileResponse
+import nemo.collections.asr as nemo_asr
+from piper.voice import PiperVoice
+
+app = FastAPI(title="ASR & TTS API")
+security = HTTPBasic()
+
+# Basic Authentication Configuration
+USERNAME = os.environ.get("API_USERNAME", "admin")
+PASSWORD = os.environ.get("API_PASSWORD", "secret")
+
+def verify_credentials(credentials: HTTPBasicCredentials = Depends(security)):
+    if not (credentials.username == USERNAME and credentials.password == PASSWORD):
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED,
+            detail="Incorrect username or password",
+            headers={"WWW-Authenticate": "Basic"},
+        )
+    return credentials
+
+# Global references for the models
+asr_model = None
+tts_voice = None
+
+@app.on_event("startup")
+async def load_models():
+    global asr_model, tts_voice
+    
+    # 1. Load and Quantize NeMo ASR
+    # Ensure you have uploaded your downloaded NeMo model to the Space with this filename
+    nemo_path = "model.nemo" 
+    if os.path.exists(nemo_path):
+        device = torch.device('cpu')
+        model = nemo_asr.models.EncDecCTCModel.restore_from(restore_path=nemo_path, map_location=device)
+        model.freeze()
+        
+        # Apply CPU Dynamic Quantization for memory reduction and speed
+        model = torch.quantization.quantize_dynamic(
+            model, {torch.nn.Linear}, dtype=torch.qint8
+        )
+        
+        model.cur_decoder = 'ctc'
+        asr_model = model
+        print("ASR Model loaded and dynamically quantized.")
+    else:
+        print("WARNING: model.nemo not found. Please upload it.")
+
+    # 2. Load Piper TTS (Nepali Chitwan)
+    tts_model_path = "chitwan.onnx"
+    if os.path.exists(tts_model_path):
+        tts_voice = PiperVoice.load(tts_model_path)
+        print("TTS Model loaded.")
+
+@app.post("/asr")
+async def transcribe(file: UploadFile = File(...), _: str = Depends(verify_credentials)):
+    if not asr_model:
+        raise HTTPException(status_code=503, detail="ASR model not loaded")
+    
+    # Save the uploaded audio temporarily
+    audio_path = f"/tmp/{file.filename}"
+    with open(audio_path, "wb") as f:
+        f.write(await file.read())
+    
+    # Run the CPU-bound ASR transcription in a thread pool
+    loop = asyncio.get_event_loop()
+    transcription = await loop.run_in_executor(
+        None, 
+        # Note: If using the AI4Bharat multilingual model, add `language_id='ne'` below
+        lambda: asr_model.transcribe(paths2audio_files=[audio_path], batch_size=1)[0] 
+    )
+    
+    os.remove(audio_path)
+    return {"text": transcription}
+
+@app.post("/tts")
+async def synthesize(text: str, _: str = Depends(verify_credentials)):
+    if not tts_voice:
+        raise HTTPException(status_code=503, detail="TTS model not loaded")
+    
+    output_path = "/tmp/output.wav"
+    
+    # Run the CPU-bound TTS synthesis in a thread pool
+    loop = asyncio.get_event_loop()
+    
+    def generate_audio():
+        with wave.open(output_path, "wb") as wav_file:
+            wav_file.setnchannels(1)
+            wav_file.setsampwidth(2)
+            wav_file.setframerate(tts_voice.config.sample_rate)
+            tts_voice.synthesize(text, wav_file)
+
+    await loop.run_in_executor(None, generate_audio)
+        
+    return FileResponse(output_path, media_type="audio/wav")