I've been looking for, and testing, various automated transcription APIs over the years and never found one that was high-quality for videos of interviews (background noise, people don't talk in full sentences, people use filler sounds). I'd love to find something usable -- and plan to try this one as well. Human transcription is laborious, slow, and expensive. I've toyed with the idea of a human clean-up pass after the automated transcription, but that's still labor intensive.

Perhaps the task you're trying to complete is just inherently laborious, slow and expensive, then?

Given that humans frequently have trouble understanding one another, it's very unlikely you'll ever encounter an API that can do 100%, or even 90% accurate transcription across environments.

---

Isn't transcription inherently an opinionated activity, as well? In terms of how to transcribe utterances, words spoken in other languages when the main language is another, etc?

Even inexpensive human transcription works pretty well though assuming good audio and recordings that aren't difficult for other reasons (strong accents, etc.) I usually get my podcasts transcribed by a service that is basically a quality control front-end for Mechanical Turk and I need to do minimal cleanup.

I'll be interested to see how this works in comparison when it's more broadly available. I'm guessing it will be sufficiently worse that it won't be worth using for this purpose. If it's cheap enough though I could see using it for other recordings that I don't normally get transcribed today.

Would you mind sharing which MT-based service you're using for podcast transcription?

CastingWords. I've been very happy with them for my purposes. They tend to even get various CamelCase product spellings and the like correct even if I forget to put them in the transcriber notes. I usually get the 6-day $1.50/min service partly because I need the turnaround and partly because, anecdotally, the quality also goes up a bit when you pay for the premium rates. I do maybe a couple dozen 20 minute podcasts a year so the costs are negligible.

Thanks for the plug, we love happy customers.

btw: We no longer use MTurk, we have our own crowd now.

Thanks. I'm dealing with hours of interviews per day.

Can you share about your use-case? I have a very similar transcription need--feel free to contact me at my email in profile

Agreed. I'm looking at transcribing some of my talks, and even if I had humans do the work I'd be doing my own editing pass. Speech is just different than writing; one needs a good mental model of both speaker and listener to make all the right editorial choices.

Automated recognition strikes me as a good place to start, though, especially for applications like search.

Yeah, I've used transcription to generate speaker notes or even give me the start of a blog post. It doesn't need to be anywhere near perfect. It's just sometimes easier/faster to get a rough dump of a presentation you're comfortable giving rather than starting with a blank sheet of paper.

Wow! Circa $1.50 hour. Whereas human transcription services are more like $45 per hour.

This kind of price will open up entirely new applications. E.g., for ~$10/day, you could transcribe every conversation you have at work. Combine that with good search and your phone becomes a supplemental, artificial memory. "I know I talked about that with somebody but I can't remember who" becomes a thing of the past.

I thought you were going to say that the human translator would use the service , add his own surcharge of $30 and still be cheaper than the competition :)

seriously this will sell like crazy to public offices and services that keep transcripts of conversations, courts of law etc.

I'm sure that will happen. Automating a first pass and saving the humans for the second, smarter pass will be a great combination. Higher quality, lower cost.

But that's only if you care about human-readable quality. I'm especially interested in the applications where we don't.

I've been looking into podcast transcription services, but I haven't really found a service build using similar APIs like this. Can anyone shed some light on the transcription business, is the transcription quality not yet good enough or why don't we see many of these services?

Shameless plug follows...!

I'd love it if you'd take a look at https://trint.com - we're designed exactly for what you're looking for. We use machine learning to deliver a highly accurate initial transcript that you can polish to perfect with our in-browser editor.

HN might be DDOSing you at the moment. Perhaps consider putting Cloudflare or similar in front of your landing page?

Microsoft and Google both have something similar. The problem is that it's IMO really not good enough for most purposes. One of these days I'll give it another try but, unless I'm just looking for a very rough cut of a talk or whatever, I would need to spend way too much time to cleanup.

Have you already looked at https://www.rev.com/ ?

Same price as google ($0.006 per 15 second increment) but more expensive that microsft ($0.004 per 15 second increment). However, this does charge by the second after 15 seconds which is nice.

I just wish they would lower the minimum threshold of 15 second intervals.

Is the quality so much better than CMU Sphinx that it's worth paying for it?

Yes. I've tried Sphinx implementations twice over the past few years. Google Cloud Speech and Nuance blew it out of the water.

It partially depends on dictionary size and the need for continuous recognition, but generally speaking there's a big divide between open source and paid services.

According to others, yes: https://news.ycombinator.com/item?id=11820837

From personal experience we can get pretty good results on our workflow from sphinx - but the amount of time it took to get those results was massive, and spread over years. It's probably worth it for most organizations to skip that and go for a vendor.

Results from Amazon are an unknown right now, but the other vendors vary a good bit and are focused on particular kinds of audio (kinds vary by: the number of speakers, the duration of the audio, the vocabulary, language).

Given Alexa's performance, this is the last thing I'd want to use.