Use faster-whisper with a streaming audio source. Includes support for asyncio.
Special thanks to JonathanFly for his initial implementation here.
This is still a work in progress, might break sometimes. Contributions welcome and appreciated!
I'd like to get this in a better state before uploading to PyPI. For now you can do
pip install git+https://github.com/pmdevita/faster-whisper-live.git
or if you're using Poetry
poetry add git+https://github.com/pmdevita/faster-whisper-live.git
LiveWhisper takes the same arguments for initialization and transcription as LiveWhisper does. However, you can pass it a file-like object, and it will stream output from it.
from faster_whisper_live import LiveWhisper
model = LiveWhisper("small", compute_type="int8_float16")
with open("whispertest.m4a", "rb") as f:
for segment in model.transcribe(f, vad_filter=True):
print(segment)Asyncio is also supported.
import asyncio
import aiofiles
from faster_whisper_live import AsyncLiveWhisper
async def main():
model = AsyncLiveWhisper("small", compute_type="int8_float16")
async with aiofiles.open("whispertest.m4a", "rb") as f:
async for segment in model.transcribe(f, vad_filter=True):
print(segment)
asyncio.run(main())If you'd like to supply your own PCM audio stream, it needs to be:
- Sample Rate: 16kHz
- Mono audio
- 16-bit Little-endian PCM (s16le)
You then need to pass decode_audio=False to transcribe.
An example ffmpeg command to generate such a file would be:
ffmpeg -i input_file.m4a -f s16le -ac 1 -ar 16000 output.pcm