Hacker News Reader

Comment by rahimnathwani | Hacker News Reader

Does anyone know how this differs in quality/speed/price from existing transcription APIs from Google, Microsoft and/or IBM?

View on HN

Haven't have a chance to test AWS Transcribe yet but I run a benchmark of IBM, Microsoft and Google transcription api weekly.

So far, IBM watson work best for english, google in other . language. For example the transcript of trump inauguration speech (16minutes 38seconds, 8371 character) give me a levenshtein distance to the official one of (smaller is better) :

- IBM : 898 (566130ms)

- Google : 1249 (301635ms)

- Microsoft : 5233 (109391ms)

Do you have some github or link as this is interesting to follow the quality of these transcription services

This is not a far comparison. IBM is the leader because they have a product that is specifically designed for this whereas the others do not.

I tried to use this, but it looks like they have to review me for access.

I've tried the IBM and Google transcription APIs on a podcast with overlapping speakers, and they turned out dadaist poetry. I've been meaning to check out other solutions.

* Google: http://mefiwiki.com/wiki/Podcast_107_Transcript_(automated)

* IBM Watson: http://mefiwiki.com/wiki/Podcast_117_Transcript_(automated)

Where's the audio? We also provide automated transcripts and I'd love to compare it.

Click the "Episode ___" link in the first paragraph, then search for "Direct mp3 download" on that page.

I'd love to see more details on what you do. I haven't found a good automated solution yet. It seems like the solutions with acceptable quality involve Mechanical Turk.

Edit: I added a Bing attempt, and I'll link the list, rather than individual eps: http://mefiwiki.com/wiki/Podcast#automated_transcripts

Here's the automated transcript from our system. I checked around a minute or so and looks like accuracy is around 80%.

https://scribie.com/transcript/c5ae0522950c421c9d939dc8a17f4...

Hmm. Is it $0.60/min for the automated solution?

Automated transcripts are free. That's the pricing for our manual transcription.

https://scribie.com/transcription/free

I've tested the google API and I'd say it's quality is similar to what I see in this example result. In other words, not great.

For me, this isn't close enough to be considered a "valid" transcription. Too much information is lost. Too much context is lost.

Truly looking forward to a day when ML / AI can do this with high degrees of accuracy, but we're not there yet.

Reply on news.ycombinator.com