whisper-live/transcriber.py

65 lines
2.0 KiB
Python

#!/usr/bin/python3
import numpy as np
import speech_recognition
import whisper
import torch
import wave
_audio_model = whisper.load_model("medium.en") # "large"
# For debugging...
# wave_out = wave.open("wave.wav", "wb")
# wave_out.setnchannels(1)
# wave_out.setframerate(16000)
# wave_out.setsampwidth(2)
class Transcriber:
def __init__(self):
self._audio_source = None
# Audio data for the current phrase.
self._current_data = b''
def set_source(self, source):
self._audio_source = source
def update(self):
if self._audio_source:
if not self._audio_source.data_queue.empty():
# We got some new data. Let's process it!
new_data = []
while not self._audio_source.data_queue.empty():
new_packet = self._audio_source.data_queue.get()
new_data.append(new_packet)
new_data_joined = b''.join(new_data)
# For debugging...
#wave_out.writeframes(new_data_joined)
self._current_data = self._current_data + new_data_joined
# Convert in-ram buffer to something the model can use
# directly without needing a temp file. Convert data from 16
# bit wide integers to floating point with a width of 32
# bits. Clamp the audio stream frequency to a PCM wavelength
# compatible default of 32768hz max.
audio_np = np.frombuffer(
self._current_data, dtype=np.int16).astype(np.float32) / 32768.0
# Run the transcription model, and extract the text.
result = _audio_model.transcribe(
audio_np, fp16=torch.cuda.is_available())
text = result['text'].strip()
print("text now: ", text)
# Automatically drop audio sources when we're finished with them.
if self._audio_source.is_done():
self._audio_source = None