How to break sentence like "hello, can you hear me?" to 2 segments by actual timestamp?

[neoyxm]

Hi , I have an audio like below: between hello and can you hear me, there is about 5s delay.

What I expect is 'hello' and 'can you hear me' are put to 2 segments:
{ "start": "00:01.680", "end": "00:02.448", "txt": "Hello" }
{ "start": "00:06.256", "end": "00:07.360", "txt": "can you hear me" }

But they are put in one segment:
{ "start": "00:00:06.256", "end": "00:00:07.360", "txt": "Hello, can you hear me?" }

Hello is merged to can you hear me segment, and segment.start (00:00:06.256) is not Hello's start time (00:01.680).

My transcribe option is : language=en, word_timestamps=True, vad_filter=True and others are default.

From the debug info from transcribing, the Silero VAD has splitted these two segments, but the result still puts them together.
[INFO]-(transcribe:240) Processing audio with duration 00:09.280
[INFO]-(transcribe:249) VAD filter removed 00:04.944 of audio
[DEBUG]-(transcribe:255) VAD filter kept the following audio segments: [00:01.136 -> 00:02.448], [00:06.256 -> 00:09.280]

Is there any options which can help to put them to 2 segments by actual timestamp?

[neoyxm]

After set word_timestamps = False, the segment.start is ok

[neoyxm]

reopen it.

[palladium123]

If I understand correctly, you want to transcribe each speech chunks separately?

If that is the case, your best bet is to split the audio separately by calling get_speech_timestamps on the audio file. This should return a List of start and end timestamps for each speech chunks, removing non-speech parts. You then want to call collect_chunks on each chunks returned from get_speech_timestamps. These two methods are in the vad.py module. Finally, call transcribe on each chunks with vad_filter=False (since you already used Silero VAD in the previous 2 steps).

Haven't tested the above approach - it's just my take looking at the transcribe method.

[neoyxm]

Thanks, I will try that.

[zacharynapier]

Did you ever find a solution?

[neoyxm]

Did you ever find a solution?

not yet

[toanhn]

I tried this approach, but sometimes I get more incorrect words or sentences in the resulting transcript than without using get_speech_timestamps.

[toanhuynhnguyen]

You can use Batched Faster-Whisper since it can transcribe an audio file to a transcript with word timestamps more accurate than sequence Faster-Whipser. Then you can separate a sentence into multiple sentences based on paused durations between words, such as 2 seconds or any N seconds you want.