11.01.2020, 23:04, "David Rowe" <da...@rowetel.com>: > 2/ With noisy speech the MELP samples appear to have removed the > interfering noise - the noise is suppressed in the output samples, and > not faithfully reproduced. This suggests the MELP implementation you > have used also has a pre-processing step to remove noise.
Yes, you were right. As noise preprocessor was a separate function in that source code, I thought that just omitting it would be sufficient. Turns out that it was not, and call to the function was also in the codec itself. I had patched out these calls and updated samples on the page. 11.01.2020, 23:04, "David Rowe" <da...@rowetel.com>: > 1/ On the noise free English samples, codec 2 at 1200/2400 is slightly > worse than MELP at the same rates. This is consistent with other tests > (e.g. academic papers using the two codecs as references). > > ... > > 3/ It's difficult for me to evaluate the Russian samples, as I don't > speak the language, but thanks for the feedback. Yes, Codec 2 was > developed on just English samples. The main issue with intelligibility that I see is that it does not improve when you give it more bitrate. When you have a good channel with [relatively] high SNR, or just UHF/VHF radio, you can't get any better output from it. Consider AMBE that does not have anything below 2 kbits, but is still heavily used on UHF/VHF just because you can afford more bandwidth there. Bitrates between, say, 1.2k and 6k are a large gap that does not have any open-source solutions. Opus can work from 6k onwards, Codec2 has good quality to bitrate ratio at 700C and below, but there is nothing in between. I had conducted a small listening test on Russian samples asking people to listen to the text and they mostly agreed that while 700C sounds much better than 450, 2400 sounds similar to 700C, despite having 3.5x bandwidth. They were unable to extract from 1200 or 2400 recording any additional words that were unintelligible on 700C. 1200 MELP encoding was perfectly intelligible, by the way. I think that distorted "robotic" sound affects intelligibility too, but it is just an assumption. MELP sounds much more naturally than Codec2. Unfortunately, I can't conduct similar test for English, since I think that it's not quite correct to measure intelligibility on non-native speakers. When I made audio samples that switches between two bitrates back and forth every 5 seconds (they can be found on samples page), I found that these 2400/3200 are virtually indistinguishable from 700C: it feels like you are just listening to a single solid recording instead of two interleaved. This, however, can be considered a good news - if 700C has almost the same intelligibility as 2400, then 2400 has a large room for improvements. This is also a good metric of a progress: your codec is now 3.5x more efficient than it was 5 years ago :) Or, considering that 3200 bitrate is also indistinguishable from 700C - even 4.6x more efficient. I'm not so much fluent in a coding theory and don't know whether knowledge gained from 700C development can be 'extrapolated' to a higher bit rates to make something like 1200B and 2400B modes, but it would be very nice to have a codec that operates in a 1k-3k bitrate range. 11.01.2020, 23:04, "David Rowe" <da...@rowetel.com>: > Hi Макс, > > Nice work on your tests! OK so this is what I hear: > > 1/ On the noise free English samples, codec 2 at 1200/2400 is slightly > worse than MELP at the same rates. This is consistent with other tests > (e.g. academic papers using the two codecs as references). > > 2/ With noisy speech the MELP samples appear to have removed the > interfering noise - the noise is suppressed in the output samples, and > not faithfully reproduced. This suggests the MELP implementation you > have used also has a pre-processing step to remove noise. > > So it's not quite an A/B comparison - the core MELP codec is being > tested on audio that has had the noise suppressed. > > With noisy input speech, the Codec 2 samples are distorted and > unpleasant to listen too, but still (as you suggest) intelligible. > > For noise suppression in Codec 2 applications we use the Speex noise > suppressor (codec2/misc/speexnoisesup.c for a command line version). > > 3/ It's difficult for me to evaluate the Russian samples, as I don't > speak the language, but thanks for the feedback. Yes, Codec 2 was > developed on just English samples. > > 4/ You are correct Codec 2 1200/2400 hasn't been actively developed in > over 5 years, recent efforts have been at lower bit rates, and LPCNet > (at 1733 bit/s). In particular to support digital voice systems for HF > radio. > > Cheers, > David > _______________________________________________ > Freetel-codec2 mailing list > Freetel-codec2@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/freetel-codec2 _______________________________________________ Freetel-codec2 mailing list Freetel-codec2@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/freetel-codec2
