miguelamendez/mica-audio8-tts-06b

Audio8 TTS Preview 0.6B — Mica Q4/Q8

Mica runtime artifacts for `Audio8/Audio8-TTS-Preview-0.6b`, pinned to revision f07040f3d151f1ba0253bfb92cb2f5dd38b44594. The model and derivatives are Apache-2.0 licensed.

The runtime configuration, conversion policy, and validation evidence are maintained in the Mica Server repository. Validated artifacts are published together in the `miguelamendez/mica-audio8-tts-06b` repository, with manifests recording their source, size, and digest.

Model facts and provenance

Audio8 is a 601,159,424-parameter DualAR TTS model, excluding its bundled 44.1-kHz codec. A 24-layer slow autoregressive transformer predicts semantic audio frames and a four-layer fast transformer predicts ten codec codebooks. It supports generation with no reference and zero-shot voice cloning.

The architecture supports up to 2,048 packed text/audio positions. The pinned checkpoint's generation configuration defaults to 512 new tokens, while the upstream example requests 1,024. Neither value is documented as a separate hard maximum output length. Training datasets, maximum training sequence, and the input/output length distribution are not disclosed.

The upstream preview recommends Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, and Spanish. A voice-cloning reference transcript must match the reference audio.

Mica artifacts and protected layers

mlx/q4/    selective MLX affine Q4, group size 64
mlx/q8/    selective MLX affine Q8, group size 64
gguf/      Q4_0 and Q8_0
vllm/      not published; structural candidates failed the production gate

The MLX conversion retains the semantic/text embeddings, acoustic codebook embeddings, fast-decoder input/output boundaries, and codec at source precision. These tensors directly cross discrete text, semantic, codebook, and waveform boundaries.

GPTQ W4A16 and W8A16 group-128 structural candidates now exercise plain and voice-reference prompts. The embeddings, codebook embeddings, fast-decoder boundaries, and waveform decoder remain at source precision. Both candidates reload and generate batch-2 WAV output, but the two-sample Hessians fall back to round-to-nearest in many layers. A full calibration and native vLLM-Omni quality run are mandatory before promotion.

For GGUF, Q40 was selected instead of Q4K because Q4_K block alignment left too many codec matrices unquantized, produced a larger resident model, and had no speed advantage.

Mica validation boundary

GGUF Q40 and Q80 passed seeded plain speech, zero-shot voice cloning, and exact Granite ASR round-trip on the Apple M4. Q4 is an accepted memory/quality compromise; Q8 was preferred in listening validation. The MLX peak-memory investigation and standardized batch suite remain open, so no MLX batch or hard memory certification is claimed here.

A matched, unseeded MLX long-form generation on the Apple M4 measured the first and final five seconds of each waveform. Q4 declined from -18.7 dB mean volume to -27.7 dB (9.0 dB) and compressed the passage into 23.78 seconds. Q8 declined from -18.3 dB to -21.0 dB (2.7 dB) and produced 34.04 seconds. Because the endpoint does not expose Audio8's sampling seed, this is a controlled prompt comparison rather than a deterministic multi-seed evaluation. Mica therefore defaults interactive TTS to Q8 and presents Q4 as a memory/quality tradeoff.

Validation records:

• —docs/validation/gguf-runtime-macos-metal.md
• —docs/validation/audio8-vllm-quantization-macos.md
• —artifacts/benchmarks/audio8-mlx-long-form-volume.json

Profile validation requirements

Audio8 profiles must state maximum packed positions, requested semantic-token output, reference-audio duration, concurrent utterances, and whether cloning is enabled. The proxy must account for the slow AR cache, fast AR fixed cache, codec workspace, and reference encoder rather than applying text-LLM KV flags. Thinking settings are invalid for this model.