speech to text to speech

fast_speech_text_speech.py

""" To use: install LLM studio (or Ollama), clone OpenVoice, run this script in the OpenVoice directory
    git clone https://github.com/myshell-ai/OpenVoice
    cd OpenVoice
    git clone https://huggingface.co/myshell-ai/OpenVoice
    cp -r OpenVoice/* .
    pip install whisper pynput pyaudio
"""

from openai import OpenAI
import time
import pyaudio
import numpy as np
import torch
import os
import re
import se_extractor
import whisper
from pynput import keyboard
from api import BaseSpeakerTTS, ToneColorConverter
from utils import split_sentences_latin

SYSTEM_MESSAGE = "You are Bob an AI assistant. KEEP YOUR RESPONSES VERY SHORT AND CONVERSATIONAL."
SPEAKER_WAV = None

llm_client = OpenAI(base_url="http://localhost:1234/v1", api_key="not-needed")

tts_en_ckpt_base = os.path.join(os.path.dirname(__file__), "checkpoints/base_speakers/EN")
tts_ckpt_converter = os.path.join(os.path.dirname(__file__), "checkpoints/converter")
device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else  "cpu"

tts_model = BaseSpeakerTTS(f'{tts_en_ckpt_base}/config.json', device=device)
tts_model.load_ckpt(f'{tts_en_ckpt_base}/checkpoint.pth')
tone_color_converter = ToneColorConverter(f'{tts_ckpt_converter}/config.json', device=device)
tone_color_converter.load_ckpt(f'{tts_ckpt_converter}/checkpoint.pth')
en_source_default_se = torch.load(f"{tts_en_ckpt_base}/en_default_se.pth").to(device)
target_se, _ = se_extractor.get_se(SPEAKER_WAV, tone_color_converter, target_dir='processed', vad=True) if SPEAKER_WAV else (None, None)
sampling_rate = tts_model.hps.data.sampling_rate
mark = tts_model.language_marks.get("english", None)

asr_model = whisper.load_model("base.en")

def play_audio(text):
    p = pyaudio.PyAudio()
    stream = p.open(format=pyaudio.paFloat32, channels=1, rate=sampling_rate, output=True)
    texts = split_sentences_latin(text)
    for t in texts:
        audio_list = []
        t = re.sub(r'([a-z])([A-Z])', r'\1 \2', t)
        t = f'[{mark}]{t}[{mark}]'
        stn_tst = tts_model.get_text(t, tts_model.hps, False)
        with torch.no_grad():
            x_tst = stn_tst.unsqueeze(0).to(tts_model.device)
            x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).to(tts_model.device)
            sid = torch.LongTensor([tts_model.hps.speakers["default"]]).to(tts_model.device)
            audio = tts_model.model.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=0.667, noise_scale_w=0.6)[0][0, 0].data.cpu().float().numpy()
            if target_se is not None:
                audio = tone_color_converter.convert_from_tensor(audio=audio, src_se=en_source_default_se, tgt_se=target_se)
            audio_list.append(audio)            
        data = tts_model.audio_numpy_concat(audio_list, sr=sampling_rate).tobytes()
        stream.write(data)
    stream.stop_stream()
    stream.close()
    p.terminate()


def record_and_transcribe_audio():
    recording = False
    def on_press(key):
        nonlocal recording
        if key == keyboard.Key.shift:
            recording = True

    def on_release(key):
        nonlocal recording
        if key == keyboard.Key.shift:
            recording = False
            return False

    listener = keyboard.Listener(
        on_press=on_press,
        on_release=on_release)
    listener.start()

    print('Press shift to record...')
    while not recording:
        time.sleep(0.1)
    print('Start recording...')

    p = pyaudio.PyAudio()
    stream = p.open(format=pyaudio.paInt16, channels=1, rate=16000, frames_per_buffer=1024, input=True)
    frames = []
    while recording:
        data = stream.read(1024, exception_on_overflow = False)
        frames.append(np.frombuffer(data, dtype=np.int16))
    print('Finished recording')

    data = np.hstack(frames, dtype=np.float32) / 32768.0 
    result = asr_model.transcribe(data)['text']
    stream.stop_stream()
    stream.close()
    p.terminate()
    return result


def conversation():
    conversation_history = [{'role': 'system', 'content': SYSTEM_MESSAGE}]
    while True:
        user_input = record_and_transcribe_audio()
        conversation_history.append({'role': 'user', 'content': user_input})

        response = llm_client.chat.completions.create(model="local-model", messages=conversation_history)
        chatbot_response = response.choices[0].message.content
        conversation_history.append({'role': 'assistant', 'content': chatbot_response})
        print(conversation_history)
        play_audio(chatbot_response)

        if len(conversation_history) > 20:
            conversation_history = conversation_history[-20:]

conversation()

Beautiful.

Does anyone have it running on a M1 MacBook Pro? Was able to get over several hurdles but stuck with the PyAudio error when its trying to stream text to speech. ||PaMacCore (AUHAL)|| Error on line 2715: err=''what'', msg=Unspecified Audio Hardware Error. OSError: [Errno -9999] Unanticipated host error. Any help is greatly appreciated!

Meanwhile some changes I had to make. 1. Change line 90 to data = np.hstack(frames).astype(np.float32) / 32768.0 2. Add Terminal to Input Monitoring within Privacy & Security Settings

@haseeb-heaven I solved this issue forcing the device to the CPU

tts_en_ckpt_base = os.path.join(os.path.dirname(__file__), "checkpoints/base_speakers/EN")
tts_ckpt_converter = os.path.join(os.path.dirname(__file__), "checkpoints/converter")
device = "cpu" # <<<

Remove the Float point warn disabling the fp16 in the transcribe section

    result = asr_model.transcribe(data, fp16=False)['text']
    stream.stop_stream()
    stream.close()
    p.terminate()

I got more error

python fast_speech_text.py
Loaded checkpoint 'checkpoints/base_speakers/EN/checkpoint.pth'
missing/unexpected keys: [] []
Loaded checkpoint 'checkpoints/converter/checkpoint.pth'
missing/unexpected keys: [] []
Press shift to record...
Start recording...
Finished recording
/opt/homebrew/Caskroom/miniforge/base/envs/hm_env/lib/python3.8/site-packages/whisper/transcribe.py:115: UserWarning: FP16 is not supported on CPU; using FP32 instead
  warnings.warn("FP16 is not supported on CPU; using FP32 instead")
[{'role': 'system', 'content': 'You are Heaven an AI assistant. KEEP YOUR RESPONSES VERY SHORT AND CONVERSATIONAL.'}, {'role': 'user', 'content': ' What are first 10 prime numbers?'}, {'role': 'assistant', 'content': ' The first ten prime numbers are 2, 3, 5, 7, 11, 13, 17, 19, 23, and 29.'}]
||PaMacCore (AUHAL)|| Error on line 2747: err=''what'', msg=Unspecified Audio Hardware Error
Traceback (most recent call last):
  File "fast_speech_text.py", line 120, in <module>
    conversation()
  File "fast_speech_text.py", line 115, in conversation
    play_audio(chatbot_response)
  File "fast_speech_text.py", line 44, in play_audio
    stream = p.open(format=pyaudio.paFloat32, channels=1, rate=sampling_rate, output=True)
  File "/opt/homebrew/Caskroom/miniforge/base/envs/hm_env/lib/python3.8/site-packages/pyaudio/__init__.py", line 639, in open
    stream = PyAudio.Stream(self, *args, **kwargs)
  File "/opt/homebrew/Caskroom/miniforge/base/envs/hm_env/lib/python3.8/site-packages/pyaudio/__init__.py", line 447, in __init__
    pa.start_stream(self._stream)
OSError: [Errno -9986] Internal PortAudio error
(hm_env) haseeb-mir@Haseebs-MacBook-Pro OpenVoice % 

I wrote a port to Spanish with facebook/tts_transformer-es-css10 model can u visit here: https://gist.github.com/iddar/9a502842ea1867d9e849cd0c03384bf2

    result = asr_model.transcribe(data, fp16=False)['text']

Thanks it worked like magic

Does anyone have it running on a M1 MacBook Pro? Was able to get over several hurdles but stuck with the PyAudio error when its trying to stream text to speech. ||PaMacCore (AUHAL)|| Error on line 2715: err=''what'', msg=Unspecified Audio Hardware Error. OSError: [Errno -9999] Unanticipated host error. Any help is greatly appreciated!

Meanwhile some changes I had to make. 1. Change line 90 to data = np.hstack(frames).astype(np.float32) / 32768.0 2. Add Terminal to Input Monitoring within Privacy & Security Settings

yes, it’s working on my device and I have the same MacBook M2 Pro and you need to make the changes as in the chat history that I have faced and I solved them now it’s working properly you need to install Python 3.8 version and then installed all the dependencies and it would run properly

someone tried to run this in ubuntu-wsl and managed to get the input from the keyboard?

someone tried to run this in ubuntu-wsl and managed to get the input from the keyboard?

try to Change the logic to use any other human input like a new line o std in

For MAC users: M1/M2 pro

1. Before installing pyaudio make sure you do - brew install portaudio
2. np.hstack(frames).astype(np.float32) - Thanks to @ajram23 above
3. Make sure you have git-lfs else checkpoints from huggingface won't be downloaded properly and you would end up getting some pkl error
4. Use requirements.txt from OpenVoice github folder to install all dependencies in one go and then additionally - pip install whisper pynput pyaudio
5. I used LLM studio with mistral as backend, make sure to start server there within local inference server

Thats lot of requirements seems like to complex project.
But it is working on my MacBook M1 now after so much complications

Yes it does on my system too. Would make changes to it , integrate speech brain probably