AAC for "mainly-speech" input?

[montython]

The typical scenario is the following:
Long (4-5 hours) TV captures mainly consisting of "talking heads" on the video side and human voice - speech (sometimes coupled with short episodes of music) on the audio side. My intention is to keep the file sizes very small, but keeping the audio clear and smooth is important.

I have a few questions on the audio side:

1) SAMPLING RATE:
I capture audio as 16-bit PCM wav at 44,100 Hz sampling rate. As expected, the frequency of the audio signals for this kind of material rarely goes above 10,000 Hz. Good quality downsampling to 22,050 yields successful results and this sampling rate seems to be optimal in line with the Nyquist–Shannon sampling theorem.

Assuming that the final result will be at 22,050 Hz sampling rate, is it preferable to capture directly at 22,050 Hz? Or should I stick to capturing at 44,100 Hz and then downsampling using SSRC for example? (Processing time is not an issue. Resampling can be a part of a long batch process which I leave alone until it finishes.)


2) CODEC CHOICE:
input: PCM Wav (mono, 22,050 Hz)
target bitrate: 16-32 kbps range

Basically I consider Real Audio COOK and Nero's AAC.

HE-AAC (SBR) doesn't seem (or should I say "sound") to make contribution in a few samples I tested myself. Considering that the input signals are generally below 10,000 Hz, can we talk about a common preference: LC-AAC vs HE-AAC?

Real COOK codec also seems to perform very well. Any comments on COOK codec vs. AAC for this particular purpose?

[Kurtnoise]

for the codec choice, you should look at speex...well tuned for speech.

[montython]

for the codec choice, you should look at speex...well tuned for speech.
The type of input I am talking about is not entirely speech, but rather "mainly" speech.

On the dedicated speech codec side, I had a look at wma speech and real audio sipro codecs as well as speex. As far as I can tell from my experience, unless you are at 1-digit bitrate levels these don't offer much compared to standard mono codecs I mentioned before.

And... Muxing these streams into mkv container is problematic...

[qyot27]

I'd use Vorbis for bitrates like that myself, as the noise floor helps immensely. Some music samples tested out as okay to my ears at 48 kb/s stereo @ 44.1kHz - it wasn't perfect, of course, but I doubt something as complex as music would be at bitrates like that; this was also 3 years ago, the precision is probably even better now, not to mention it's mostly speech being looked at as per this thread's topic. There were a couple music samples which still got slaughtered by that much compression, but they were rather Industrial Rock-sounding - most of the other samples were fine, and they were a mix of several genres.

I would expect, based on that, that dropping the frequency, using mono, and being mainly speech would make that perfectly suitable. And it muxes into MKV easily.

[montython]

I would expect, based on that, that dropping the frequency, using mono, and being mainly speech would make that perfectly suitable.
In fact, I made some reading on the subject. From what I have gathered, I understand that almost any audio codec (including lame mp3) will produce quite good results @32 kbit/s assuming mono input at 22,050 sampling rate. (I posted my gatherings at another topic.)

I also made a blind listening test myself and my ears verified this. I included lame mp3, real cook, wma, nero aac as well as the reference samples @ 44,100 and 22,050 sampling rates. It was really really difficult to say the differences. The funny thing is: at the end of the day I discovered that I gave an higher score to real cook compared to the reference @44,100.:)

In conclusion, I guess the codec choice is not very critical for this type of input assuming a bitrate level of 32 kbit/s. I will prefer to stay at 32 kbit/s level. Nevertheless, I will repeat the test for 16 kbit/s level for experimental purposes. Probably the differences will be more significant at this lower bitrate. This time I will include Vorbis and see what my ears tell me.

Thank you for your interest.