app.py

from collections import deque
import os
import threading
import time
import av
import numpy as np
import streamlit as st
from streamlit_webrtc import WebRtcMode, webrtc_streamer
import pydub
# import av
# import cv2
from sample_utils.turn import get_ice_servers
import json
from typing import List

from vosk import SetLogLevel, Model, KaldiRecognizer
SetLogLevel(-1) # mutes vosk verbosity

from dotenv import load_dotenv
load_dotenv()

system_one = {
    "audio_bit_rate": 16000,
    # "audio_bit_rate": 32000,
    # "audio_bit_rate": 48000,
}


playing = st.checkbox("Playing", value=True)

def load_vosk (model='small'):
    # load vosk model
    # get path of current file
    current_file_path = os.path.abspath(__file__)
    current_directory = os.path.dirname(current_file_path)
    _path = os.path.join(current_directory, 'models', 'vosk', model)
    model_voice = Model(_path)
    recognizer = KaldiRecognizer(model_voice, system_one['audio_bit_rate'])
    return recognizer

vask = load_vosk()

def handle_audio_frame(frame):
    # if self.vosk.AcceptWaveform(data):
    pass


def do_work(data: bytearray) -> tuple[str, bool]:
    text = ''
    speaker_finished = False
    if vask.AcceptWaveform(data):
        result = vask.Result()
        result_json = json.loads(result)
        text = result_json['text']
        speaker_finished = True
    else:
        result = vask.PartialResult()
        result_json = json.loads(result)
        text = result_json['partial']
    return text, speaker_finished


frames_deque_lock = threading.Lock()
frames_deque: deque = deque([])

async def queued_audio_frames_callback(
            frames: List[av.AudioFrame],
        ) -> av.AudioFrame:
    with frames_deque_lock:
        frames_deque.extend(frames)

    # create frames to be returned.
    new_frames = []
    for frame in frames:
        input_array = frame.to_ndarray()
        new_frame = av.AudioFrame.from_ndarray(
            np.zeros(input_array.shape, dtype=input_array.dtype),
            layout=frame.layout.name,
        )
        new_frame.sample_rate = frame.sample_rate
        new_frames.append(new_frame)
    
    # TODO: replace with the audio we want to send to the other side.

    return new_frames

webrtc_ctx = webrtc_streamer(
    key="charles",
    desired_playing_state=playing,
    # audio_receiver_size=4096,
    # audio_frame_callback=process_audio,
    queued_audio_frames_callback=queued_audio_frames_callback,
    mode=WebRtcMode.SENDRECV,
    rtc_configuration={"iceServers": get_ice_servers()},
    async_processing=True,
)

system_one_audio_status = st.empty()

if not webrtc_ctx.state.playing:
    exit

system_one_audio_status.write("Initializing...")
system_one_audio_output = st.empty()
system_one_audio_history = []
system_one_audio_history_output = st.empty()


sound_chunk = pydub.AudioSegment.empty()
while True:
    if webrtc_ctx.state.playing:
        audio_frames = []
        with frames_deque_lock:
            while len(frames_deque) > 0:
                frame = frames_deque.popleft()
                audio_frames.append(frame)

        if len(audio_frames) == 0:
            time.sleep(0.1)
            system_one_audio_status.write("No frame arrived.")
            continue

        system_one_audio_status.write("Running. Say something!")

        for audio_frame in audio_frames:
            sound = pydub.AudioSegment(
                data=audio_frame.to_ndarray().tobytes(),
                sample_width=audio_frame.format.bytes,
                frame_rate=audio_frame.sample_rate,
                channels=len(audio_frame.layout.channels),
            )
            sound = sound.set_channels(1)
            sound = sound.set_frame_rate(system_one['audio_bit_rate'])
            sound_chunk += sound

        if len(sound_chunk) > 0:
            buffer = np.array(sound_chunk.get_array_of_samples())
            text, speaker_finished = do_work(buffer.tobytes())            
            system_one_audio_output.markdown(f"**System 1 Audio:** {text}")
            if speaker_finished and len(text) > 0:
                system_one_audio_history.append(text)
                if len(system_one_audio_history) > 10:
                    system_one_audio_history = system_one_audio_history[-10:]
                table_content = "| System 1 Audio History |\n| --- |\n"
                table_content += "\n".join([f"| {item} |" for item in reversed(system_one_audio_history)])
                system_one_audio_history_output.markdown(table_content)
            sound_chunk = pydub.AudioSegment.empty()

    else:
        system_one_audio_status.write("Stopped.")
        break