The spectral quantizer gets only the bits that are left after all side information is accounted for:

The encoder runs an 8-iteration bisection search to find the global gain index ggind (0–255) such that the resulting quantized spectrum uses approximately nbits_spec bits.

Gain offset ggoff is computed once per frame from the bitrate and sampling rate:

The energy E[k] (in dB) of blocks of 4 MDCT coefficients is computed for the bisection. The bisection starts at ggind=255 and tests 8 halvings, adjusting ggind up/down based on whether the estimated bit count exceeds the adjusted budget nbits_spec’ (which includes a carry-over offset from the previous frame’s actual vs. estimated bit usage).

A minimum gain ggmin is enforced to prevent the quantized spectrum from exceeding the 16-bit range [−32768, 32767]:

The global gain is first dequantized:

Each coefficient is then quantized using a dead-zone quantizer with dead-zone offset ±0.375:

The ±0.375 dead zone (instead of standard ±0.5) slightly biases toward smaller-magnitude quantized values, producing slightly more zeros which compress better. Coefficients smaller than 0.375×gg in magnitude are quantized to 0.

After quantization, the encoder estimates how many bits it would take to arithmetically encode Xq. This estimation drives two decisions: (1) truncation of trailing zeros, and (2) lsbMode flag. The algorithm walks through spectral 2-tuples (pairs of coefficients), simulating the arithmetic coder context to estimate bit cost using the ac_spec tables from Section 3.7.7.

Two bitrate flags are set first:

rateFlag shifts the arithmetic coder context table offset for higher-bitrate content. modeFlag enables the high-bitrate LSB mode where the LSBs of large coefficients are sent separately in the residual section.

If the estimated bit count exceeds nbits_spec, trailing zero coefficients are cut:

The lsbMode flag is set to 1 when modeFlag=1 and even after truncation the bit count still exceeds nbits_spec. In lsbMode, the LSBs (least significant bits) of high-value coefficients are stripped from the arithmetic coder stream and stored separately in the backward-written side information area, providing fine-grained bit savings.

After bit estimation, if the actual bit count is significantly different from the target, ggind is adjusted by ±1 or +2 and the spectrum is re-quantized once. The adjustment thresholds delta and delta2 are adaptive — they are larger at high bitrates (where you can afford to be more generous) and smaller at low bitrates. The three tables t1, t2, t3 govern the threshold transitions:

The adjustment runs at most once per frame. If re-quantization is done, the original nbits_est is preserved for the next frame’s offset computation.

Residual coding uses any remaining bits (beyond what arithmetic coding needs) to refine non-zero coefficients. It is only active when lsbMode = 0.

For each non-zero Xq(k), one bit is stored indicating whether the original Xf(k) was above or below the quantization midpoint:

nbits_residual_max = nbits_spec − nbits_trunc + 4
k = 0; nbits_residual = 0;
while (k < NE && nbits_residual < nbits_residual_max):
    if Xq[k] != 0:
        res_bits[nbits_residual] = (Xf[k] >= Xq[k]*gg) ? 1 : 0
        nbits_residual++
    k++

At the decoder, these residual bits shift each non-zero quantized value by ±0.1875 or ±0.3125, improving accuracy without full re-quantization.

Spectral coefficients that are quantized to zero are not transmitted — but at the decoder, they need to be filled with something to avoid silence in those frequency bins (which sounds harsh). The decoder uses noise filling, substituting random noise at a level controlled by the transmitted noise factor FNF.

The relevant zero coefficients are those within the active bandwidth and surrounded by other zeros within ±NFwidth bins (i.e., isolated zeros, not zeros at the spectral edge). The mean level of the TNS-filtered coefficients in this mask is computed and normalized by the global gain:

LNF is quantized to 8 levels:

FNF = 7 means LNF ≈ 0 (nothing to fill). FNF = 0 means LNF ≈ 0.5 (strong noise fill level). FNF is transmitted as 3 bits in the side information.