Best compressed audio?

[Steam Ranger]

I'm trying to fit a 3 minute song onto a floppy disk, the best format I've discovered so far is AMR, which has low quality but wonderful amounts of compression. I have several questions:

1. Does anyone know a better format than AMR for a stock 68000? (More quality with same size)
2. Where can I obtain Asm/C code to decode the better format on the Amiga? (Real time)
3. If AMR is the best, can someone provide me with Asm/C decoder (Takes in a memory buffer pointer, outputs a pointer to a new memory buffer with decoded sound)

Thanks in advance.

[britelite]

What quality is your input audio and what's the desired output frequency and bit depth? Mono, stereo?

I assume you've done some benchmarks with AMR, what kind of results did you get regarding size and quality?

[Steam Ranger]

Originally Posted by britelite

What quality is your input audio and what's the desired output frequency and bit depth? Mono, stereo?

I assume you've done some benchmarks with AMR, what kind of results did you get regarding size and quality?

With the highest quality narrowband AMR I can fit the Audio on a DD disk 3 times, I don't know how fast/slow the compression is, but I think it might be a bit slow decompressing? I can't find and way to test it.

AMR narrowband is locked at 8000Hz, and broad band is non-distributable.

I'll settle for 8000Hz minimum, any number of bits I can't hear the difference. It must be able to decode in real time, preferable stream able off of disk.

[hooverphonique]

Ask the guy who did this: https://www.pouet.net/prod.php?which=65736

[Steam Ranger]

That sounds like tracker music. I'm converting an MP3 song to fit on a floppy disk, and to be usable by a stock A500. It has lyrics, so I can't just convert it to tracker music.

[britelite]

Originally Posted by Steam Ranger

That sounds like tracker music. I'm converting an MP3 song to fit on a floppy disk, and to be usable by a stock A500. It has lyrics, so I can't just convert it to tracker music.

It's not tracker music though

[stevelord]

Second half has lyrics.

[Steam Ranger]

Oh right, didn't watch that far.

[no9]

There is another one [ Show youtube player ]

[SoDa7]

Originally Posted by Steam Ranger

I'm trying to fit a 3 minute song onto a floppy disk, the best format I've discovered so far is AMR, which has low quality but wonderful amounts of compression. I have several questions:

1. Does anyone know a better format than AMR for a stock 68000? (More quality with same size)
2. Where can I obtain Asm/C code to decode the better format on the Amiga? (Real time)
3. If AMR is the best, can someone provide me with Asm/C decoder (Takes in a memory buffer pointer, outputs a pointer to a new memory buffer with decoded sound)

Thanks in advance.

Honestly, I'm not sure that AMR is the best format for music - as AMR is a format optimized for voice (and so, used for phone calls mostly)

I think that instead, for an unexpanded AMIGA (AMR might be too much for a 68000 afaik), you could try 8SVX at 8000hz mono. it would be something like 64kbps per second of PCM audio (on an AMIGA formatted floppy is approx 1 minute and 40 seconds of audio)

[Don_Adan]

For Amiga sample it can be 50% compression, i dont know if AMR did it better.

[pandy71]

I would check some open source codecs as Speex and/or Opus (SILK) they should outperform closed and proprietary AMR - but honestly doubt if you can do something useful on plain MC68000 @7MHz with all normal codecs - maybe some heavily stripped and simplified MPEG 1 Layer I. Or perhaps some INT decoders as Tremor for Vorbis or Helix for MP3 but still >>onslaught<< seem to be best in terms quality/compression/CPU utilization.
ADPCM may be sane approach - giving compression 4 or more...

[smack]

If the AMR you mean is this one https://en.wikipedia.org/wiki/Adapti...te_audio_codec
then I think this decoder can't run in realtime on the 7 MHz 68000.
Also, that's a narrow-band speech codec so I expect the music quality to be seriously degraded.

IMO a reasonable approach can be:
- reduce the sample rate as much as possible, while still getting acceptable audio quality (for example 8 kHz or lower)
- use an ADPCM codec to compress the audio data to 4 bits per sample (50% reduction)

You can find a ready-to-use ADPCM codec implementation by Mikael Kalms here:
https://github.com/Kalmalyzer/adpcm-68k

example:
3 minutes, 8 kHz mono, 4-bit ADPCM would fit on a single floppy disk
180s * 8000/s * 4b = 5.760.000b = 703 KiB

[Steam Ranger]

Thanks for all the help! I've managed to fit it onto the disk as ADCMP, however I was hoping for more than 8Khz, so I'm thinking I'll devise my own simple compression and build it into my program and it will run while it's streaming from disk.

[meynaf]

Alternatively you may try XPK lib. Its SQSH method is efficient on samples and fast enough for 68000.

[Steam Ranger]

Originally Posted by meynaf

Alternatively you may try XPK lib. Its SQSH method is efficient on samples and fast enough for 68000.

Is XPK a generic archive or audio format?

[meynaf]

Originally Posted by Steam Ranger

Is XPK a generic archive or audio format?

It's a compression library with sub-libraries - some of which being specialized for audio.
http://aminet.net/package/util/pack/xpk_User

[Karlos]

If you want low CPU usage and the ability to pretty much play directly without decompressing the whole thing (or large blocks first), basic 4-bit delta encoding is a good trade off. For the same storage requirements of 8kHz you could use 16kHz. There are different ways to encode the 16 values, I've experimented with quite a few, but you have to be able to encode negative and positive deltas. Also you want to use a nonlinear encoding ideally. Delta value distribution favours small values compared to the original signal. I've also experimented with a zeroless scheme for this (where you can't represent an unchanging value). It's not as limiting for real audio signals as it sounds, but it isn't particularly good on more basic waveforms for reasons that are hopefully self-evident.

For maximum simplicity, you might choose deltas 0, 1, 2, 4, 16, 32, 64, 127, -1, -2, -4, -8, -64. The lack of symmetry is why I experimented with zeroless encoding.

Reading a block of delta values and outputting the sample stream is trivial. You just keep track of the current value, get the next 4-bit delta index, look it up in your encoding table above and add it on. That's the next output sample.

[paraj]

Originally Posted by Karlos

If you want low CPU usage and the ability to pretty much play directly without decompressing the whole thing (or large blocks first), basic 4-bit delta encoding is a good trade off. For the same storage requirements of 8kHz you could use 16kHz. There are different ways to encode the 16 values, I've experimented with quite a few, but you have to be able to encode negative and positive deltas. Also you want to use a nonlinear encoding ideally. Delta value distribution favours small values compared to the original signal. I've also experimented with a zeroless scheme for this (where you can't represent an unchanging value). It's not as limiting for real audio signals as it sounds, but it isn't particularly good on more basic waveforms for reasons that are hopefully self-evident.

For maximum simplicity, you might choose deltas 0, 1, 2, 4, 16, 32, 64, 127, -1, -2, -4, -8, -64. The lack of symmetry is why I experimented with zeroless encoding.

Reading a block of delta values and outputting the sample stream is trivial. You just keep track of the current value, get the next 4-bit delta index, look it up in your encoding table above and add it on. That's the next output sample.

I think this a good approach for getting 50%, but you want to change the "127" to "-128" in that table, haven't checked if other changes are better. Did a super-quick check of what difference that makes (attached: wagner1.wav is your table, wagner2.wav is using same with 127 changed to -128). (You can hear the original at https://www.lemonamiga.com/games/details.php?id=701 by selecting 8svx.wagner)

[Karlos]

Originally Posted by paraj

I think this a good approach for getting 50%, but you want to change the "127" to "-128" in that table, haven't checked if other changes are better. Did a super-quick check of what difference that makes (attached: wagner1.wav is your table, wagner2.wav is using same with 127 changed to -128). (You can hear the original at https://www.lemonamiga.com/games/details.php?id=701 by selecting 8svx.wagner)

The table was to be illustrative rather than a real world example. To be clear, for real audio, the chances of getting a full scale d deflection in a delta signal is diminishingly small. The delta is the differential of the signal. Think about what the real signal must look like to have such a delta.