[opus] Antw: Re: Antw: Re: Antw: Re: Possible bug in Opus 1.3

Ulrich Windl Ulrich.Windl at rz.uni-regensburg.de
Tue Nov 6 06:54:52 UTC 2018

>>> Emily Bowman <silverbacknet at gmail.com> schrieb am 05.11.2018 um 20:46 in
<CAGSVXPR6t8uHJFqDCT-1pn9otP_7ypPxgRXRasgZERunsAe0fA at mail.gmail.com>:
> On Mon, Nov 5, 2018 at 11:01 AM Jan Stary <hans at stare.cz> wrote:
> Attached I send the spectrogram (vic SoX) of the first 20 seconds
>> for the wav file and the opus file. Indeed, there is extra noise
>> for the low frequencies, but somewhere around -100 dB.
>>         Jan
> That might be entirely due to SoX treating it as a 16-bit file, which it is
> not; -100dB is almost exactly the limitation of 16-bit. All Opus files are
> infinite-precision, and they'll encode the input at whatever precision is
> fed to them, but they do have a silence-detection mechanism which defaults
> to 16-bit in opusenc. SoX is either reading that value and erroneously
> assuming that's that internal precision, or just falling back on a 16-bit
> default because it has no information, but it would be more proper to
> _always_ decode to 24-bit to eliminate that broadband noise low bit depths
> create (except in the case of hardware limitations).
> As you've found, Opus is always 48kHz, never more, never less. Its
> resampler is very accurate, and should never introduce noise.

That's an interesting point: Why is that so? Couldn't the file size (and computing effort on encoding/decoding) be reduced when allowing different sample rates? Especially when the source has lower sampling rate, I wonder whether that isn't just a waste of bits...

> My speakers and headphones definitely have issues with the sweep, so it's
> hard to isolate any differences. (I do love it though, it really gets my
> floor shaking; I'll keep it around for testing purposes.) The only
> remaining issue is the way that Opus doesn't even come close to respecting
> the requested bitrate for this sample. For instance, encoding it at 130
> gives me a file of 210kbps. Over a wide corpus of music, I've noticed
> libopus almost invariably overshoots the bitrate, rather than averaging out
> close to it; only speech is consistently at or below it.

Maybe there's some bug still undiscovered. Would you agree that Opus should encode a pure sinus tone efficiently? Or would it really prefer to encode white noise?


More information about the opus mailing list