looking for efficient speech codec Options

[chrizoo]

Hi. I have quite a few speech recordings (talkradio, lectures, notes to self, etc.) which I need to encode. I have done some tests with HE-AAC V2 (Nero AAC codec / 1.5.3.0) and am quite pleased with the quality/filesize ratio.

I then tried speex (speex-1.2beta1/$). When comparing to HE-AACv2 files of the same size, speex failed horribly for me during VLC 1.1.11 playback, suffering from major artefacts and lower overall quality. Not quite what I expected. (Unless VLC decoding is broken or the speex version too old ?)

1. How does speex fare against HE-AACv2 for you ?
   
2. Is there any other substantially more efficient speech codec ?
   
3. What software player (Windows) can you recommend ?
   
4. What considerations should I not overlook if I want to keep the files for life (I'm still young :-) ?

many thanks

[jensend]

I'll address point 4 first: Bluntly, if you want to keep the files for life, you probably want to keep them in a lossless format. Space is cheap, and formats, encoders, and player support will change over the years. Having a lossless copy means you can always re-encode into whatever the lossy format of the day is.

1. Speex should beat HE-AACv2 at some low bitrates, but its main advantage over HE-AAC and even Vorbis is not quality but latency. It will lose to HE-AAC and Vorbis at moderate or high bitrates.

2. First, the bad news. There's nothing substantially more efficient than HE-AAC with really solid player support right now.

The good news is that this is changing. The Opus codec is already vastly superior to both Speex and HE-AAC for speech, and there's still room for more improvements in the reference encoder. It is just about to release its 1.0 version. Not much player support yet but it will be there quite soon.

3. I find myself just using VLC since it's pretty handy for a lot of stuff.


Just checking- have you been resampling your files before passing them to your HE-AAC encoder? General-purpose codecs like HE-AAC or Vorbis, since they've been tuned for music, will avoid resampling so as to preserve the quality of music's high frequencies, but for speech recordings those higher frequencies are just noise, and removing them allows the encoder to spend more bits on things that matter. Depending on your hearing and preferences, the sample rate sweet spot for straight speech could be anywhere from 12kHz to 24kHz.