Chatterbox-Turbo-MLX-Community-6bit

A tested MLX Chatterbox Turbo checkpoint published by groxaxo, with loading files, generic default-voice sample audio, precision provenance and measured comparison evidence. This repository is one of the three quantized candidates selected by measured controlled decoder fidelity, alongside community affine 6-bit, oQ6 and oQ5. The unquantized reference and oQ4 are comparison baselines.

Source and precision

Source: mlx-community/chatterbox-turbo-6bit, revision afbfd1f8a32f2bc1f32e0fb6b7663c04327e8ff1. Original model: ResembleAI/chatterbox-turbo; MLX conversion credit: mlx-community and mlx-audio.

Unchanged community affine 6-bit checkpoint, republished from the pinned source revision. It has a different tensor inventory from the oQ derivatives. Its 134 shared stored convolution tensors match the original after dtype conversion.

Weights SHA-256: 50737677896b0e35a7526662c127841040830b826ef8a2d017bddcea6f23e4ef. Checkpoint size: 534.6 MiB (model.safetensors only). The full repository includes tokenizer/conditioning files and evaluation assets. The exact stored dtype inventory, source/config/conditioning hashes and measured results are in provenance.json. The bit label is not a measured uniform precision for every tensor.

Load and generate on Apple Silicon

These weights were tested with MLX 0.32.2, mlx-audio 0.4.8, and mlx-lm 0.32.0 on an Apple M5 Mac with 24 GiB unified memory. The repository includes conds.safetensors for the bundled default voice.

pip install 'mlx==0.32.2' 'mlx-audio==0.4.8' 'mlx-lm==0.32.0'
python -m mlx_audio.tts.generate --model groxaxo/Chatterbox-Turbo-MLX-Community-6bit --text "Hello, this is Chatterbox Turbo speaking locally." --play

from mlx_audio.tts.utils import load_model
import mlx.core as mx
import numpy as np
from scipy.io import wavfile

model = load_model("groxaxo/Chatterbox-Turbo-MLX-Community-6bit")
parts = []
for result in model.generate(text="Hello, this is Chatterbox Turbo speaking locally."):
    mx.eval(result.audio)
    parts.append(np.asarray(result.audio).reshape(-1))
audio = np.concatenate(parts)
wavfile.write("output.wav", result.sample_rate, (np.clip(audio, -1, 1) * 32767).astype(np.int16))

Listen to the generic bundled-voice sample. For reference/candidate A/B samples and all metrics see evaluation/README.md and evaluation/index.html. No private voice recording or model weights from other variants are included.

Measured comparison

Profile	Checkpoint MiB	Warm median s	RTF	Controlled speaker cosine	Whole-model speaker cosine	Controlled log-mel difference dB
Original · FP32 stored weights	2847.7	2.861	0.431	1.00000	1.00000	0.000
Community affine 6-bit	534.6	1.535	0.254	0.99908	0.96628	1.308
oQ6	643.1	1.489	0.247	0.99618	0.96603	1.839
oQ5	547.4	1.445	0.235	0.97661	0.94626	3.013
oQ4	463.9	1.389	0.226	0.93667	0.90098	5.130

Timing is the median of five warm repetitions of one sentence; output durations differ and RTF normalizes generation time by generated duration. Lower RTF means faster generation. This is a battery-powered desktop rerun with Low Power Mode off. The earlier run used Low Power Mode on; speed differences cannot be attributed to power mode alone. oQ6's separate recheck was 1.547 s / RTF 0.256. Small gaps among quantized candidates do not establish a universal speed ranking.

Controlled decoder metrics reuse original speech tokens, shared bundled conditioning and decoder RNG seed 9182. They evaluate decoder fidelity conditional on those tokens. Full generation uses synthesis seed 20261009 and shared sampling settings; pacing and sampled tokens can diverge. The in-family speaker encoder is shared across candidates. Cosine and spectral error are not percentages of audible quality, MOS or independent speaker verification.

All 30/30 rerun comparison WAVs matched the earlier experiment's hashes exactly. The experiment has three English sentences, one synthesis seed and one bundled voice. No human preference/MOS, independent speaker encoder, multilingual, cloned-voice or long-form certification is claimed. These measurements support community 6-bit as a provisional speed/fidelity candidate and oQ6 as the strongest oQ fidelity candidate on this small corpus.

For this model, peak MLX allocation was 1.913 GiB, peak process RSS 1.176 GiB and load time 0.917 s. These are full-process peaks, not total system RAM or steady per-request memory. Complete-generation strict pooled WER was 3.92%. Fresh ASR used CPU INT8 Parakeet; the earlier run used CUDA INT8 of the same named model. Four transcripts changed despite identical WAVs. ASR errors do not establish naturalness or isolate recognition versus synthesis omissions.

Evidence and reproducibility

• Comprehensive report, summary CSV, per-clip CSV, similarity JSON.
• Weight audit, fresh checkpoint inventory, all clip hash comparisons.
• Timing plot, fidelity plot, storage trade-off plot.
• Frozen packages, generation/measurement scripts, Experimentos source.

Model weights are unchanged since their tested SHA-256 fingerprints. SHA256SUMS covers the published repository's payload files (excluding itself). To host the listening page, download evaluation/ and run python evaluation/scripts/serve.py --bind 127.0.0.1 --port 8789.

Attribution and licenses

The MLX source repositories declare Apache-2.0; this derivative retains that metadata and includes LICENSE. Resemble AI's original Chatterbox Turbo model card and software declare MIT; the original copyright notice is preserved in LICENSE-UPSTREAM-MIT. NOTICE and the source model card preserve provenance. groxaxo claims the local quantization/evaluation work and does not claim authorship of the original model or community conversion.