How to find the pair of "audio-clip" & "text"?
Hi,
Thank you for providing such a large amount of speech data.
I'm trying to use it to train a speech recognition system.
After downloaded datasets from "ivrit-ai/audio-transcripts" & "ivrit-ai/audio-vad", I was confused by their meta data.
I cannot find the correct pair of "audio-clip" & "text".
Could you kindly help me on this?
Hello Joseph,
The two mentioned datasets may not be the best source for what you are after.
ivrit-ai/audio-vad -> audio only, sliced using "vad" to small (avg ~5s) audio clips, without continuity (gaps are omitted).
ivrit-ai/audio-transcripts -> machine generated transcripts of the above, to be improved by humans (but before it was improved).
Those are derived from the (now legacy) dataset: ivrit-ai/audio-base which includes the source audio files + metdata.
The "audio-base" has been replaced recently by ivrit-ai/audio-v2 which has more content.
Soon we will release machine generated transcriptions of this dataset, using better models (WER's are better across the board).
Still, if you want high quality audio+transcripts which are human generated take a look at:
ivrit-ai/crowd-transcribe-v5 (the short segments from ivrit-ai/audio-transcripts, reviewed and fixed by humans, about 300h)8800 audio length, About ~4700h good quality with transcripts, word level timestamps)
ivrit-ai/crowd-recital - Human recorded transcripts of Wikipedia articles (high quality - about ~50h, word level timestamps)
ivrit-ai/knesset-plenums - Long audio recordings + human generated transcripts in varying qualities (
Hopefully, The dataset cards would provide enough data for your usage.