Compatibility with faster-whisper

[jbkkd]

faster-whisper is a much faster implementation of Whisper, done using ctranslate2.

Is there a path forward to make this great library compatible with faster-whisper?

[jianfch]

see #75

[chislon]

Using the example from #152 https://github.com/jianfch/stable-ts/blob/main/examples/non-whisper.ipynb one can convert the result from faster_whisper.

(Below might not work as-is but some variation of this)

from faster_whisper import WhisperModel
from stable_whisper import WhisperResult
from tqdm import tqdm

MODEL_SIZE = "medium"
LANGUAGE = "zh"
model = WhisperModel(model_size_or_path = MODEL_SIZE)

path="test.mp3"

segments, info = model.transcribe(
	audio=path,
	language=LANGUAGE,
	condition_on_previous_text=False,
	vad_filter=True,
	word_timestamps=True
)

segments_list = []
total_duration = round(info.duration, 2)
with tqdm(total=total_duration, unit=" seconds") as pbar:
	for segment in segments:
		segment_duration = segment.end - segment.start
		pbar.update(segment_duration)
		segments_list.append(segment)

result = []
for segment in segments_list:
	word_dicts = []
	for word in segment.words:
		word_dict = {
			'start': word.start,
			'end': word.end,
			'word': word.word
		}
		word_dicts.append(word_dict)
	result.append(word_dicts)    

result = WhisperResult(result)
result = (
	result.split_by_punctuation([('.', ' '), '。', '?', '？', ',', '，'])
	.split_by_gap(.5)
	.merge_by_gap(.3, max_words=3)
	.split_by_punctuation([('.', ' '), '。', '?', '？'])
	.split_by_length(max_words=11)
	.merge_by_punctuation(["'"])
)