What six bits preserved in Chatterbox Turbo

Five checkpoints were compared in two cohorts. Complete generations allow speech-token choices and pacing to change. Controlled decoder runs reuse the original tokens and conditioning, allowing a narrower acoustic comparison.

Separate token choices from decoding

The supposedly fp16 source actually stored FP32 weights. Inspecting the tensors corrected that label before drawing storage conclusions.

What the controls showed

Community 6-bit used 534.6 MiB, versus 2,847.7 MiB for the original. Controlled speaker cosine was 0.99908 and log-mel difference 1.308 dB. oQ6 measured 0.99618 and 1.839 dB; oQ5 and oQ4 moved farther away. Full-generation speaker cosine was almost tied between community 6-bit and oQ6, near 0.966.

The oQ files contained 318 convolution tensors that had been quantized and subsequently stored in fp16. Dequantizing them made a loadable floating representation; it did not restore the original values.

Timing needs its conditions

The first run used battery power with Low Power Mode enabled. Community 6-bit, oQ5 and oQ4 occupied a similar observed speed band, while oQ6 varied substantially. Those observations do not identify a universal speed winner.

Keep community 6-bit as the provisional candidate for this default English workload. More text, multiple voices, independent encoders and blinded listening are the next useful evidence.

Evidence and publication boundary

This is an edited account of the recorded experiment. The linked archive is private; the methods, aggregate findings and limits are included above. Reference recordings, raw conversations, authentication material and workstation configuration are excluded.