Full Version: metallic at low bitrates
I've encoded some audio book at 32kbs, 22khz, mono. All is great but this matallic sound (especially at beter quality source !!) really irritating me. How can I remove it? Some cut off or sth?
use LAME with '--lowpass x' switch
where 'x' is a integer number for cutting frequency. metallic sound is heard because of small bitrate/big freq-band, or because of the bad encoder - nothing but a big harmonic distorsion
where 'x' is a integer number for cutting frequency. metallic sound is heard because of small bitrate/big freq-band, or because of the bad encoder - nothing but a big harmonic distorsion
[FONT=Arial][FONT=Arial][FONT=Arial]
Any suggestion? cut at 15, lower, higer? I know that I have to do some test and check it myself, but now I don't have source at home
so now I ask you (about your opinions) and I'll try it later...
QUOTE(Med0 @ Mar 12 2003 - 03:10 AM)
use LAME with '--lowpass x' switch
where 'x' is a integer number for cutting frequency.
where 'x' is a integer number for cutting frequency.
Any suggestion? cut at 15, lower, higer? I know that I have to do some test and check it myself, but now I don't have source at home
so now I ask you (about your opinions) and I'll try it later...
1. I've got some audiobook encodings too (32 ... 64kbps) and noticed that metallic sound / harshness of "s" depends on the player (on some hardwareplayers it's almost gone, on others (of the cheaper sort) it's annoying. As you didn't say what you used for playback - maybe it's a decoder problem (aditionally).
2. If you resample to 22.05 kHz the lowpass needs to be below 11.025 kHz, otherwise it'd be meaningless, so you could try e.g. --lowpass 10.
3. You could try some settings mentioned in one of these threads.
2. If you resample to 22.05 kHz the lowpass needs to be below 11.025 kHz, otherwise it'd be meaningless, so you could try e.g. --lowpass 10.
3. You could try some settings mentioned in one of these threads.
QUOTE(lucjansz @ Mar 13 2003 - 01:27 PM)
Any suggestion? cut at 15, lower, higer?
Lower... frequencies of human speech go roughly from 0.3 to 8 KHz, with the highest energy around 2 to 3 KHz, the most sensitive hearing range. Try --lowpass 6 to --lowpass 4.
QUOTE(CiTay @ Mar 13 2003 - 01:55 PM)
Lower... frequencies of human speech go roughly from 0.3 to 8 KHz [...]
8 kHz ? That's what the telecom people want you to believe..
enough for understanding the message, but really not so great quality for sibilants or female voice 
Edit: Oh. Didn't see that the bitrate was 32kbps.. sorry.
And for voice telecom (POTS), they actually give you about 4kHz frequency response (8kHz being the sample rate, half that being the Nyquist limit), so 8 kHz (16 kHz sample rate) is quite a lot better.
For .MP3 a rule of thumb is that it usually sounds very crummy and metallic or slushy at around one-bit-per-sample, but gets fairly OK at about 1.5 to 2 bits per sample.
If you're using lowpass to achieve a reduction in information to be encoded you can get an equivalent rule of thumb. I'd say that for mono sound, the lowpass frequency in kHz should be about one-quarter to one-third of the bitrate in kbps to give reasonable quality, but that's only a rule of thumb and you might need to cut further after listening (particularly to sibilant sounds) on test excerpts.
If you aim for 32 kbps, I'd guess that about 11 kHz lowpass might do, but that's basically 22.05 kHz sampling rate, where you say you've got problems, so you might try about 8 kHz lowpass (effectively 2 bits per sample, or 4 bits per Hz of bandwidth) which is most efficiently handled at 16kHz sampling rate (the next sampling rate above double the lowpass frequency).
ABR mode is also likely to help (or even better, VBR if you lowpass it and specify the sample rate to bring it to around 32 kbps on average)
Assuming ABR, something like:
Assuming VBR, something like:
For VBR, play around with the quality setting, -V 4 (try -V 2, -V 6 etc.) and home in on roughly 32 kbps for your typical samples, and/or play with the resample frequency (16, 22, 24, 32 kHz) and lowpass frequency.
With VBR the best approach may be to start with --resample 22.05 and no lowpass setting then home in on the maximum value -V n for which the metallic sound is OK (and perhaps one lower for a bit of safety). Then go about adjusting the sampling rate and/or lowpass until you achieve around 32 kbps on average (if that's what you're aiming for).
VBR lets you choose the quality first and let the bitrate take care of itself, which sounds like what's important to you, but to get the bitrate reasonably low, you need to force it to give up bandwidth to save bits by reducing the sampling rate (which indirectly lowpasses) and maybe the lowpass to fine tune it.
Note that below 16 kHz sampling rate, MPEG 2.5 layer 3 is used. This isn't strictly in the MP3 standard, so might not be supported by hardware MP3 players that support MPEG 1 layer 3 and MPEG 2 layer 3.
I've found for Ogg Vorbis that a particular quality setting (i.e. VBR) is pretty good for speech (see my sig), and the bitrate then depends on the bandwidth of my lowpass or the sampling rate, as well as the complexity of my material.
For .MP3 a rule of thumb is that it usually sounds very crummy and metallic or slushy at around one-bit-per-sample, but gets fairly OK at about 1.5 to 2 bits per sample.
If you're using lowpass to achieve a reduction in information to be encoded you can get an equivalent rule of thumb. I'd say that for mono sound, the lowpass frequency in kHz should be about one-quarter to one-third of the bitrate in kbps to give reasonable quality, but that's only a rule of thumb and you might need to cut further after listening (particularly to sibilant sounds) on test excerpts.
If you aim for 32 kbps, I'd guess that about 11 kHz lowpass might do, but that's basically 22.05 kHz sampling rate, where you say you've got problems, so you might try about 8 kHz lowpass (effectively 2 bits per sample, or 4 bits per Hz of bandwidth) which is most efficiently handled at 16kHz sampling rate (the next sampling rate above double the lowpass frequency).
ABR mode is also likely to help (or even better, VBR if you lowpass it and specify the sample rate to bring it to around 32 kbps on average)
Assuming ABR, something like:
CODE
lame -m m --resample 16 --lowpass 7.7 --abr 32
Assuming VBR, something like:
CODE
lame -m m --resample 16 --lowpass 7.7 --V 4
For VBR, play around with the quality setting, -V 4 (try -V 2, -V 6 etc.) and home in on roughly 32 kbps for your typical samples, and/or play with the resample frequency (16, 22, 24, 32 kHz) and lowpass frequency.
With VBR the best approach may be to start with --resample 22.05 and no lowpass setting then home in on the maximum value -V n for which the metallic sound is OK (and perhaps one lower for a bit of safety). Then go about adjusting the sampling rate and/or lowpass until you achieve around 32 kbps on average (if that's what you're aiming for).
VBR lets you choose the quality first and let the bitrate take care of itself, which sounds like what's important to you, but to get the bitrate reasonably low, you need to force it to give up bandwidth to save bits by reducing the sampling rate (which indirectly lowpasses) and maybe the lowpass to fine tune it.
Note that below 16 kHz sampling rate, MPEG 2.5 layer 3 is used. This isn't strictly in the MP3 standard, so might not be supported by hardware MP3 players that support MPEG 1 layer 3 and MPEG 2 layer 3.
I've found for Ogg Vorbis that a particular quality setting (i.e. VBR) is pretty good for speech (see my sig), and the bitrate then depends on the bandwidth of my lowpass or the sampling rate, as well as the complexity of my material.
How does MPC --standard fare with speech, bitrate-wise? Does it get a relatively low bitrate? I'd do something like that myself, but I lack any decent digital speech. 
This is a "lo-fi" version of our main content. To view the full version with more information, formatting and images, please click here.