Holding speech tokens fixed made the Q8 advantage clearer
This follow-up applied the Turbo method to Multilingual FP32, Q8 and Q6. Thirty-six full generations and thirty-six controlled decoder outputs formed the primary cohort; warm timing and FP32 self-repeats were retained separately.
Key finding
Q8 had lower controlled decoder error than Q6 on the paired English and Spanish clips.
Edited finding from the pinned experiment record. The report and figure metadata preserve the original workload and limits.
The boundary
Three phrases per language and two seeds. Decoder fidelity excludes token-generation changes and cannot establish naturalness.
Data
Reviewed aggregate results with explicit units, sources and qualification limits. Different cohorts remain separate.
Q8 stayed closer under the fixed-token control
Three English + three Spanish phrases · two seeds · same FP32 conditioning
- Q8 G640.332
- Q6 G641.001
log-mel RMS difference (dB)
Decoder fidelity is not naturalness. Whole generation can change tokens, duration and omissions.
View data & source
| Configuration | Value |
|---|---|
| Q8 G64 | 0.332 |
| Q6 G64 | 1.001 |
Origin: reviewed report table. Revision 7e6fdbe3b1d1.
Write-ups
Report chatterbox-multilingual/fidelity-2026-10-10/README.md
A fresh protocol for a different model
This follow-up applied the Turbo method to Multilingual FP32, Q8 and Q6. Thirty-six full generations and thirty-six controlled decoder outputs formed the primary cohort; warm timing and FP32 self-repeats were retained separately.
Same tokens, closer decoder
Mean controlled log-mel difference was 0.332138 dB for Q8 and 1.000980 for Q6. Controlled speaker cosine was 0.999778 versus 0.998462. Full-generation pooled strict WER was 18/216 for FP32, 29/216 for Q8 and 43/216 for Q6.
Those totals preserve omissions and cap-reaching outliers. Q8 omitted a second sentence in two English clips; one Q6 Spanish output hit the 400-token ceiling with substantially changed pacing. Good averages did not erase these failures.
Check the numerical floor
The original decoder’s self-repeats differed at roughly float32 rounding scale. None of twelve controls were PCM16 byte-identical, although their maximum float amplitude delta was only 1.19e-7. The final protocol measured that floor instead of requiring an unrealistic equality condition.
Q8 remained the selected fidelity default for this small corpus. No human listening, emotion or broad speaker-identity claim follows from the cosine values.
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.
More in Speech & codecs
MOSS on a 3090: Q2 failed, AWQ4 stayed closer
Every profile began with the pinned BF16 generator. The comparison reused fixed reference codes and an FP32 decoder for the main quality measurements, keeping codec changes separate from generator quantization.
AWQ4 stayed close to the BF16 reference on WavLM and character errors; the sensitivity Q2 candidate was rejected.
Drama3: BF16 decoded the same codes about three times faster
The experiment replayed saved Spanish and English codes representing 21.5481 and 15.8824 seconds of audio. It compared FP32, BF16, mixed BF16 and two INT8 decoder variants on one RTX 3090.
BF16 reduced median decoder time from 390.07 to 129.54 ms in Spanish and from 292.42 to 96.88 ms in English.
Keeping FP32 and moving Supertonic to Triton
Rather than reduce the CPU graph precision again, this follow-up used NVIDIA Triton’s ONNX backend with CUDA execution and the unchanged FP32 graphs. The paired comparison retained both speed and acoustic checks.
The paired GPU test cut median first-audio time to about 0.35 seconds while preserving the measured ASR results.