.TTA File Format – Prediction and channels

Hybrid filter

Each channel has its own filter of 8 taps, with weights qm, a delay line dl and step values dx. The shift depends on the sample width, and the rounding value is 1 << (shift − 1).

FFmpeg’s table has a fourth entry, 12, for 32-bit samples, which its decoder does not accept.

For each input value:

1. Adapt the weights by the sign of the previous input. If it was negative, qm[i] −= dx[i]; if positive, qm[i] += dx[i]; if 0, nothing.
2. Sum dl[i] × qm[i] over the 8 taps, adding the rounding value, and shift right.
3. Shift dx[0..3] and dl[0..3] down by one, and set dx[4..7] from the signs of dl[4..7], with magnitudes 1, 2, 2 and 4.
4. Keep the input as the sign to use next time.
5. The output is the input plus the shifted sum.
6. Set dl[7] to the output, and dl[6..4] to differences that cascade through the delay line.

All of the state starts at 0 at the start of a frame, apart from the weights in encrypted files.

Fixed predictor

After the filter, each value adds a first-order prediction from the channel’s last output:

Rebuilding the channels

TTA stores differences between neighbouring channels, so once every channel of a sample position has been decoded, it rebuilds them in reverse. The last channel gains half the value of the one before it, and each earlier channel is then the next channel minus its own value:

last channel += value of the channel before ÷ 2
for each earlier channel, from the last:
    channel = next channel − channel

Output

Passwords

A file with audio format 2 is encrypted with a password. FFmpeg computes a CRC-64 of the password with the polynomial 0x42F0E1EBA9EA3693, and uses its 8 bytes, sign-extended, as the filter’s initial weights.