normativecurrent normative draft · unfrozen

ALA1 audio bitstream — Draft Generation 1

Normative decoded semantics and bitstream syntax for the Draft Generation 1 ALA1 audio codec.

Warning

Draft status: experimental, unfrozen, and subject to incompatible change while the software is 0.x. V1/ALV1/ALA1 refer to Draft Generation 1, not a stable 1.0 compatibility promise.

Status: normative draft for Draft Generation 1. It defines the current reference-decoder contract, but compatibility is not frozen.

ALA1 is a reversible-lifting transform audio codec with scalar quantization. qstep=1 is mathematically lossless. V1 Baseline interoperability requires 48 kHz mono or stereo; the packet syntax itself permits 1–8 channels and sample rates from 8 kHz through 384 kHz.

1. Packet

bytes 0..3    ASCII "ALA1"
byte  4       channels, 1..8
byte  5       flags
bytes 6..7    qstep u16, >=1
bytes 8..11   sample_rate u32
bytes 12..13  sample_frames u16, 1..4096
bytes 14..15  reserved, MUST be zero
bytes 16..19  entropy_len u32
bytes 20..    entropy block

Flag bit 0 is reversible mid/side coupling and is legal only for two channels. All other flag bits MUST be zero. The entropy block is defined by the common V1 entropy specification.

No priming or pre-skip is required by ALA1.

2. Channel preparation

Input is signed 16-bit PCM represented as mathematical integers.

If stereo mid/side is enabled, for every sample pair (L,R) compute:

side = R - L
mid  = L + (side >> 1)
plane0 = mid
plane1 = side

>> 1 is arithmetic right shift, equivalent to floor division by 2 for signed two’s-complement integers.

3. Blocking and token stream

Block length is 64 samples per channel. The final block is zero-padded to 64 before transform; only the declared sample_frames are returned after decoding.

The decompressed entropy token stream begins with canonical unsigned varint block_count, which MUST equal ceil(sample_frames/64).

Then, channel-major and block-major, each transformed block is sparse-coded:

uvarint nonzero_count          // 0..64
repeat nonzero_count times:
    uvarint coefficient_index  // 0..63, strictly increasing
    svarint quantized_value

A coefficient value MUST have absolute value ⇐2,000,000.

4. Reversible lifting Haar transform

For a 64-value vector, repeatedly transform the current prefix length len, initially 64 and halved until 1. For each pair (a,b):

d = b - a
s = a + (d >> 1)

Store all s values in the first half of the current prefix and all d values in the second half.

The inverse begins at prefix length 2 and doubles through 64. For each (s,d):

a = s - (d >> 1)
b = d + a

and interleave (a,b).

This lifting transform is exactly reversible over integers absent arithmetic overflow.

5. Quantization

Let base quantizer q = qstep.

For coefficient index i:

if q == 1, divisor is 1 for every coefficient;
otherwise, divisor is q for i < 4, 2q for 4 ⇐ i < 16, and 3q for i >= 16.

Encoder quantization rounds to nearest with halves away from zero:

nonnegative: (v + d/2)/d;
negative: -((-v + d/2)/d).

Decoder dequantization is exact multiplication by the divisor. Consequently qstep=1 has no quantization and is lossless.

6. Stereo inverse

If mid/side was enabled:

L = mid - (side >> 1)
R = side + L

Every reconstructed channel sample MUST fit signed 16-bit PCM; otherwise the packet is invalid.

7. Arithmetic validity

All lifting and stereo inverse additions/subtractions are mathematically exact operations but a V1 decoder MUST reject a packet if an intermediate cannot be represented in signed 32-bit arithmetic or if a final PCM sample cannot be represented in signed 16-bit arithmetic.