\Integrate Sarvam AI STT engine"

app.py

@@ -1,5 +1,6 @@
 
 import os
+import asyncio
 import logging
 import tempfile
 import time
@@ -12,6 +13,8 @@ from pydantic import BaseModel
 from faster_whisper import WhisperModel
 import azure.cognitiveservices.speech as speechsdk
 import subprocess
+import base64
+from sarvamai import AsyncSarvamAI
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
@@ -206,6 +209,92 @@ async def transcribe_azure(
             if p and os.path.exists(p):
                 os.unlink(p)
 
+@app.post("/api/transcribe/sarvam", response_model=TranscriptionResponse)
+async def transcribe_sarvam(
+    file: UploadFile = File(...),
+    language: Optional[str] = Form(None)
+):
+    """Transcribes audio using Sarvam AI (Saaras v3)"""
+    key = os.getenv("SARVAM_API_KEY")
+    if not key:
+        raise HTTPException(status_code=500, detail="Sarvam API key not configured on server.")
+
+    # Map language codes to Sarvam format (e.g., 'te' -> 'te-IN')
+    # Sarvam supports: hi-IN, bn-IN, kn-IN, ml-IN, mr-IN, or-IN, pa-IN, ta-IN, te-IN, gu-IN, en-IN
+    lang_map = {
+        "te": "te-IN",
+        "hi": "hi-IN",
+        "en": "en-IN",
+        "ta": "ta-IN",
+        "kn": "kn-IN",
+        "ml": "ml-IN",
+        "bn": "bn-IN",
+        "mr": "mr-IN",
+        "gu": "gu-IN",
+        "pa": "pa-IN"
+    }
+    
+    sarvam_lang = lang_map.get(language, "en-IN")
+    
+    temp_path = None
+    try:
+        validate_audio_file(file)
+        content = await file.read()
+        
+        # Convert to base64 as required by the streaming SDK example
+        audio_data = base64.b64encode(content).decode("utf-8")
+        
+        client = AsyncSarvamAI(api_subscription_key=key)
+        
+        logger.info(f"Sarvam Transcribing {file.filename} | Lang: {sarvam_lang}")
+        
+        async with client.speech_to_text_streaming.connect(
+            model="saaras:v3",
+            mode="transcribe",
+            language_code=sarvam_lang,
+            high_vad_sensitivity=True
+        ) as ws:
+            await ws.transcribe(audio=audio_data)
+            # Finish transcription
+            await ws.transcribe(audio="")
+            
+            # Collect results
+            full_text = ""
+            while True:
+                try:
+                    # Sarvam streaming returns partial components
+                    response = await asyncio.wait_for(ws.recv(), timeout=10.0)
+                    if not response:
+                        break
+                    
+                    # Log the response structure for debugging
+                    logger.debug(f"Sarvam Response: {response}")
+                    
+                    # The response is usually a dict or object with transcript
+                    # Based on typical streaming SDKs, it might have 'transcript' or similar
+                    if hasattr(response, 'transcript'):
+                        full_text += response.transcript
+                    elif isinstance(response, dict) and 'transcript' in response:
+                        full_text += response['transcript']
+                    
+                    # If it's a final response, break
+                    if getattr(response, 'is_final', False) or (isinstance(response, dict) and response.get('is_final')):
+                        break
+                        
+                except asyncio.TimeoutError:
+                    break
+            
+            return TranscriptionResponse(
+                text=full_text.strip() or "Transcription failed or returned empty.",
+                language=language or "unknown",
+                confidence=1.0,
+                duration=0.0
+            )
+
+    except Exception as e:
+        logger.error(f"Sarvam STT Error: {e}")
+        raise HTTPException(status_code=500, detail=str(e))
+
 if __name__ == "__main__":
     import uvicorn
     uvicorn.run(app, host="0.0.0.0", port=8000)

requirements.txt

@@ -9,3 +9,4 @@ ffmpeg-python
 # Force rebuild trigger for large-v3 model download
 numpy
 azure-cognitiveservices-speech
+sarvamai