circletreef/stable-ts-con

Stable timestamps and confidence score for words of OpenAI's Whisper outputs down to word-level.

Stabilizing Timestamps and confidence scores for Whisper

Description

This script modifies methods of Whisper's model to gain access to the predicted timestamp tokens of each word (token) without needing additional inference. It also stabilizes the timestamps down to the word (token) level to ensure chronology. Additionally, it can suppress gaps in speech for more accurate timestamps.

Confidence score is a log_softmax used for sum_logprobs also.

F.log_softmax(logits.float(), dim=-1)

TODO

• Add function to stabilize with multiple inferences

Dependency

• Whisper

Setup

Option 1: Install Whisper+stable-ts-con (one line)

pip install git+https://github.com/Anoncheg1/stable-ts-con.git

Option 2 (without confidence): Install Whisper (repo) and stable-ts (PyPI) separately

1. Install Whisper
2. Check if Whisper is installed correctly by running a quick test

import whisper
model = whisper.load_model('base')
assert model.transcribe('audio.mp3').get('segments')

3. Install stable-ts

pip install stable-ts

Executing script

from stable_whisper import load_model

model = load_model('base')
# modified model should run just like the regular model but with additional hyperparameters and extra data in results
results = model.transcribe('audio.mp3')
stab_segments = results['segments']
first_segment_word_timestamps = stab_segments[0]['whole_word_timestamps']

# or to get token timestamps that adhere more to the top prediction
from stable_whisper import stabilize_timestamps
stab_segments = stabilize_timestamps(results, top_focus=True)

Generate .srt with stable timestamps

# word-level
from stable_whisper import results_to_word_srt
# after you get results from modified model
# this treats a word timestamp as end time of the word
# and combines words if their timestamps overlap
results_to_word_srt(results, 'audio.srt')  # combine_compound=True will merge words with no prepended space

# sentence/phrase-level
from stable_whisper import results_to_sentence_srt
# after you get results from modified model
results_to_sentence_srt(results, 'audio.srt')
# below is from large model default settings

jfk_phrases.mp4

# sentence/phrase-level & word-level
from stable_whisper import results_to_sentence_word_ass
# after you get results from modified model
results_to_sentence_word_ass(results, 'audio.ass')
# below is from large model default settings

jfk_phrases_words.mp4

Additional Info

• Since the sentence/segment-level timestamps are predicted directly, they are always more accurate and precise than word/token-level timestamps.
• Although timestamps are chronological, they can still be off sync depending on the model and audio.
• The unstable_word_timestamps are left in the results, so you can possibly find better way to utilize them.

License

This project is licensed under the MIT License - see the LICENSE file for details

Acknowledgments

Includes slight modification of the original work: Whisper