Which is best sound format for IBM Speech to Text?

Question

IBM advises using Opus sound format for audio submitted to its Watson Speech to Text service. The idea being that Opus is designed specifically for speech.

Otherwise, it says you will get better quality transcription when submitting audio in flac format than in mp3 format. The latter has the obvious advantage of its small size. There is after all a 100Mb limit for file submissions. So you weigh the balance of your needs. That all makes sense so far.

But looking at conversions done on a source WAV file, the Opus file is size is comparable with mp3.

Downsampling a 366Mb wav file to 8k sample rate (one of two sample rates advised for using the service), created a wav file of 66.4Mb. Converting that to flac, wav and opus produced flac: 43.6Mb; mp3: 6.2Mb; opus: 9.8Mb.

So is opus really the best choice for getting the most accurate transcription? And how can that be when it is so small compared to flac?

Answer 1

Opus is designed to efficiently encode speech. The details are explained in the linked wiki article, but just to give you a gist, consider that human vocalisation range is rather limited, roughly from 80 to 260 Hz. On the other hand, or hearing range is far greater, up to 20000 Hz. Whereas music encoders (like mp3) have to work roughly within our hearing range, voice-specialised encoders (like Opus) can focus on what matters to efficiently encode human voice, with no interest what lies significantly above our vocalisation range. That I hope provides some intuition why Opus is so efficient.

Is it the best? It's somewhat opinionated, but yes, I think it's among the best choices out there. To cite after Wikipedia, Opus replaces both Vorbis and Speex for new applications, and several blind listening tests have ranked it higher-quality than any other standard audio format at any given bitrate.