23 Sep 2022
Paper: Robust Speech Recognition via Large Scale Weak Supervision
Model Card (?):
3rd Party Resources
Getting Started
install
pip3 install git+https://github/openai/whisper.git
usage
Working 1st try:
- in less than 3mns
- with a .m4a file (requiring no conversion to .wav)
- with 4 lines of code (!)
- with very good accuracy
🤯 😁
import whisper
model = whisper.load_model("base")
result = model.transcribe('test/1.m4a')
print(f'\n{result["text"]}\n')
What is missing for voice transcripts, is voice commands recognition, eg "comma", "new line", "full stop", etc..
I can add those in my script though.
!projects/transcribee
25 Sep 2022
It also detects language automatically, and pretty accurately so far!
response structure
{ 'language': 'en',
'segments': [ { 'avg_logprob': float,
'compression_ratio': float,
'end': float,
'id': int,
'no_speech_prob': float,
'seek': int,
'start': float,
'temperature': float,
'text': str,
'tokens': [ int,
int,
int,
...,
]}],
'text': str # returned transcript
}
example:
file = 'test/test_from_watch.m4a'
import whisper
model = whisper.load_model("base")
result = model.transcribe(file)
pp.pprint(result)
outputs:
UserWarning: FP16 is not supported on CPU; using FP32 instead
warnings.warn("FP16 is not supported on CPU; using FP32 instead")
Detected language: english
{ 'language': 'en',
'segments': [ { 'avg_logprob': -0.4222838251214278,
'compression_ratio': 0.9866666666666667,
'end': 11.200000000000001,
'id': 0,
'no_speech_prob': 0.11099651455879211,
'seek': 0,
'start': 0.0,
'temperature': 0.0,
'text': ' This is a test of a transcript in English '
'recorded from my watch fullstop.',
'tokens': [ 50364,
639,
307,
257,
1500,
295,
257,
24444,
294,
3669,
8287,
490,
452,
1159,
1577,
13559,
13,
50924]}],
'text': ' This is a test of a transcript in English recorded from my watch '
'fullstop.'}
Note: need to figure out what the UserWarning is about 🤔
26 Sep 2022
working really well in implementation now in use of Transcribee: !projects/transcribee
testing for meeting recordings
would be great to have a timed transcript, similar to a subtitle file.
found:
# Write a transcript to a file in SRT format.
# Example usage:
from pathlib import Path
from whisper.utils import write_srt
result = transcribe(model, audio_path, temperature=temperature, **args)
# save SRT
audio_basename = Path(audio_path).stem
with open(Path(output_dir) / (audio_basename + ".srt"), "w", encoding="utf-8") as srt:
write_srt(result["segments"], file=srt)
from: https://github/openai/whisper/blob/main/whisper/utils.py
Newest version includes a write_srt function.
Need to update
Testing models
29 Sep 2022
Only used the base model so far.
Testing medium.en now.
1.42Gb download.
Getting now a ValueError(f"This model doesn't have language tokens so it can't perform lang id")
testing long file
01 Oct 2022
With Tesla's AI Day online since today, I wanted to get a transcript of it.
Video file downloaded from Youtube is 14Gb.
Extracted audio-only !helpers/extract-audio-from-video resulting in a 197Mb .m4a file.
Good news:
Whisper transcribed it in 133 minutes with settings as:
model = whisper.load_model("medium")
response = model.transcribe(file_path,language='english')
Bad news:
152,288 characters on 7 lines!
Hope to see improvements re layout and speaker recognition in future updates.
Will check tomorrow for accuracy & SRT implemenentation.
04 Oct 2022
SRT generation implemented with:
from whisper.utils import write_srt
model = whisper.load_model("medium")
response = model.transcribe(file_path,language='english')
# text = response["text"]
# language = response["language"]
srt = response["segments"]
parent_folder = os.path.abspath(os.path.join(file_path, os.pardir))
# write SRT
with open(f"{parent_folder}/{uid}.srt", 'w', encoding="utf-8") as srt_file:
write_srt(srt, file=srt_file)
Again, works surprisingly well 😁
See example with a 3.5h video (transcribed in 2h) at !interests/tesla-ai-day-2022
Command line
Can be used as
whisper audio_or_video_file.m4a --language en --model medium
Mac Apps
15 Mar 2024
Available now in Setapp: superwhisper - testing now.
Otherwise this seems to be the best?
