sentence segmentation in Wav2vec2

[MisakaMikoto-o]

how can i use ctc to segment sentence in Wav2vec2? the exsample just segment words

[lumaku]

In the Wav2vec2 example given in README.md, the variable transcripts is a list of two sentences:

transcripts = ["A MAN SAID TO THE UNIVERSE", "SIR I EXIST"]

This then should give you separated alignments for each sentence:

print(align_with_transcript(audio,transcripts))
# [{'text': 'A MAN SAID TO THE UNIVERSE', 'start': 0.08124999999999993, 'end': 2.034375, 'conf': 0.0}, 
#  {'text': 'SIR I EXIST', 'start': 2.3260775862068965, 'end': 4.078771551724138, 'conf': 0.0}]

The example also gives you the word timestamps.

print(get_word_timestamps(audio))
# [{'text': 'a', 'start': 0.08124999999999993, 'end': 0.5912715517241378, 'conf': 0.9999501323699951}, 
# {'text': 'man', 'start': 0.5912715517241378, 'end': 0.9219827586206896, 'conf': 0.9409108982174931}, 
# {'text': 'said', 'start': 0.9219827586206896, 'end': 1.2326508620689656, 'conf': 0.7700278702302796}, 
# {'text': 'to', 'start': 1.2326508620689656, 'end': 1.3529094827586206, 'conf': 0.5094435178226225}, 
# {'text': 'the', 'start': 1.3529094827586206, 'end': 1.4831896551724135, 'conf': 0.4580493446392211}, 
# {'text': 'universe', 'start': 1.4831896551724135, 'end': 2.034375, 'conf': 0.9285054256219009}, 
# {'text': 'sir', 'start': 2.3260775862068965, 'end': 3.036530172413793, 'conf': 0.0}, 
# {'text': 'i', 'start': 3.036530172413793, 'end': 3.347198275862069, 'conf': 0.7995760873559864}, 
# {'text': 'exist', 'start': 3.347198275862069, 'end': 4.078771551724138, 'conf': 0.0}]

However, these timestamps do not represent alignments in the sense of start-ending of a word, but rather the most probably timestep of occurrence.

Other than that - If your transcripts is a list of words, the example should give you word segments? If you modified the example code, I suggest to check the datatype of transcripts.