5 ms·
Regarding non-streaming transcription: It doesn't seem like the return includes timings, which would limit its usefulness in transcription for accessibility (wh
by spdustin 10y ago
Regarding non-streaming transcription: It doesn't seem like the return includes timings, which would limit its usefulness in transcription for accessibility (where, for example, a post-processing step on our side would consume the Cloud Speech API and build SRT subtitle files for class recordings)
Is there a product / service / library out there that could help align a known-good transcript with the original audio (creating the timings) if the Cloud Speech API doesn't plan to return timings?
- nolite 10y agoWhy would you need the Speech API if you already have a known-good transcript?
- spdustin 10y agoI'm sorry if I was unclear. If I used the Cloud Speech API to produce the transcript, is there another library/service/etc that would align it to the original audio again.
- taf2 10y agoWatson does this, VoiceBase does this and I believe HP and MS do this... it's very common for speech API's to provide time offset information for each word.
- nmcfarl 10y agoSpeechmatics sells this product. My company, CastingWords, uses it to align our human produced transcripts, and make SRTs. It works well.
- danso 10y agoYeah that was a bummer to see in the alpha and I was holding out that they'd expose that data at release date (as IBM Watson has), but maybe they want to limit the ability for competitors to easily appropriate the sound-to-transcribed-word data.
- spdustin 10y agoThanks for mentioning Watson; I keep meaning to go back to it and see how limits have improved.
- skoocda 10y agoI'm currently developing a system along those lines. Not publicly available quite yet, but should be soon. Bug fixing at the moment.
- nshm 10y agoYou can use Speechmatics directly