Use `libopus@32kbps/48kHz/mono` for best size to quality ratio

[veganomy]

libopus at 32 kbps 48kHz mono channel provides the smallest file size with best voice quality at that bitrate.

I'm currently using https://github.com/stsaz/phiola with the above configuration and being getting the smallest filesize with best audio transparency.

[m3thm4th]

There is not a perfect fit for everyone. For example you chose 48 kHz which is overkill for voice recording, and want it inside a measly 32 kbps file.

User should be able to decide parameters:

• Sampling rate: 8 - 16 - 24 - 32 - 48 kHz
• Audio depth: 16 - 24 - 32 bits
• Bitrate: 32 - 48 - 64 - 96 - 128 kbps

I agree Opus is the best solution for encoding, especially for the use case of this app. It is FOSS and well supported by Android, there's no need for proprietary and inefficient formats like WAV.

[veganomy]

48 kHz is overkill for voice recording, and want it inside a measly 32 kbps file.

@m3thm4th No it's not. libpus v1.2 reaches voice transparency around 12 kbps. Opus can actually handle 48Khz very well even at 16 kbps. I have tried and tested all bitrate / bit depth / sample size scenarios that you listed.

Going below 32 kbps (upto 16 kbps), you would start noticing pipe like sounds. Going below 16 bit depth, you would lose the sound stage. Going below 48Khz starts giving muffling voice with damped frequencies. Based on this, I've found that 32kbps/16bit/48khz/mono produces the most voice transparency at the lowest file size.

And when I went from 32 - 64 - 96 kbps, I actually didn't found a major quality upgrade. Except you'll only start noticing slight difference on wide sound stage headphones or high particle home theatre. For continuous 24x7 voice recording, I believe 32k would be a perfect sweet spot. It's even better setting to mono channel, as you're recording on a phone mic. It'll also improve frequencies at 32k.

I've recorded 24 hours with above config, which give around 500mB file size with near perfect voice quality.

[m3thm4th]

You didn't do your homework right, but it doesn't matter because my point is that users should have the ability to decide which parameters suit them best.

[dotancohen]

Note that voice transcription is usually done at 16 kHz. This is the recommended value by Whisper and Google Chirp, and is the value used for the LibreSpeech library.

Regarding bit rate, I have not tested with voice but I typically use 128 kbps - I can actually hear the difference between 128 and 196kpbs MP3 files (or could when I was younger). Not tested with Opus (which I use for voice).

[dotancohen]

This issue seems to be a dupe of #13

[veganomy]

opus can handle very high sampling rates even at very low bitrates. That's why it is considered to achieve speech transparency quicker than any other lossy codecs.

That's exactly what we want. Extremely clear audio quality at low file sizes. That is how it is designed.

128kbps for opus is pretty much overkill. Opus reaches transparency at around 32 kbps & full band stereo quality at 64 kbps. Going further 96-128 kbps makes sense when there are multichannel dolby formats.