Spaces:
Running
Running
| # qa.py | |
| import os | |
| import json | |
| import tempfile | |
| import streamlit as st | |
| from streamlit_webrtc import webrtc_streamer, WebRtcMode, RTCConfiguration, AudioProcessorBase | |
| from utils import generate_audio_mp3, call_groq_api_for_qa | |
| import av | |
| import pydub | |
| import wave | |
| import numpy as np | |
| # For streaming from the mic, we need some RTC configuration | |
| RTC_CONFIGURATION = RTCConfiguration( | |
| {"iceServers": [{"urls": ["stun:stun.l.google.com:19302"]}]} | |
| ) | |
| class AudioBufferProcessor(AudioProcessorBase): | |
| """ | |
| A custom audio processor that accumulates raw audio frames in memory. | |
| When the user stops, we can finalize them into a single WAV for STT. | |
| """ | |
| def __init__(self) -> None: | |
| self.frames = [] | |
| def recv_audio(self, frame: av.AudioFrame) -> av.AudioFrame: | |
| # Convert the audio frame to a pydub AudioSegment | |
| pcm = frame.to_ndarray() | |
| # The shape is (channels, samples) | |
| # We'll assume single channel or handle the first channel | |
| if pcm.ndim == 2 and pcm.shape[0] > 1: | |
| # If stereo, just take the first channel for STT | |
| pcm = pcm[0, :] | |
| sample_rate = frame.sample_rate | |
| samples = pcm.astype(np.int16).tobytes() | |
| segment = pydub.AudioSegment( | |
| data=samples, | |
| sample_width=2, # int16 | |
| frame_rate=sample_rate, | |
| channels=1 | |
| ) | |
| self.frames.append(segment) | |
| return frame | |
| def finalize_wav(self) -> str: | |
| """ | |
| Once the user stops recording, combine frames into a single WAV file. | |
| Returns path to the wav file. | |
| """ | |
| if not self.frames: | |
| return "" | |
| combined = sum(self.frames) | |
| with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_wav: | |
| combined.export(tmp_wav.name, format="wav") | |
| return tmp_wav.name | |
| def handle_qa_exchange(conversation_so_far: str, user_question: str) -> (bytes, str): | |
| """ | |
| 1) Build system prompt from conversation_so_far + user_question | |
| 2) Call the LLM to get short JSON | |
| 3) TTS the answer | |
| 4) Return (audio_bytes, answer_text) | |
| """ | |
| system_prompt = f""" | |
| You are John, the guest speaker. The user is asking a follow-up question. | |
| Conversation so far: | |
| {conversation_so_far} | |
| New user question: | |
| {user_question} | |
| Please respond in JSON with keys "speaker" and "text", e.g.: | |
| {{ "speaker": "John", "text": "Sure, here's my answer..." }} | |
| """ | |
| raw_json_response = call_groq_api_for_qa(system_prompt) | |
| response_dict = json.loads(raw_json_response) | |
| answer_text = response_dict.get("text", "") | |
| speaker = response_dict.get("speaker", "John") | |
| if not answer_text.strip(): | |
| return (None, "") | |
| # TTS | |
| audio_file_path = generate_audio_mp3(answer_text, "John") | |
| with open(audio_file_path, "rb") as f: | |
| audio_bytes = f.read() | |
| return (audio_bytes, answer_text) | |