appautomaton/dots-tts-mlx

dots.tts SOAR and MeanFlow — MLX

This repository contains four self-contained dots.tts inference artifacts for `mlx-speech`. They run the full text-to-waveform pipeline in MLX on Apple Silicon without a PyTorch, Transformers, or mlx-lm inference dependency.

Variants

The short aliases select int8 because both quantized artifacts passed the local release gate described below. Base and int8 artifacts share one repository but load independently; mlx-speech downloads only the selected subtree and this root model card.

Architecture and precision

dots.tts uses text scheduling and a Qwen2.5 contextual trunk to generate continuous latent speech patches autoregressively. A causal semantic encoder feeds generated patches back into Qwen. SOAR uses flow matching with classifier-free guidance; MeanFlow uses a distilled solver without a separate runtime guidance branch. CAM++ supplies reference-speaker conditioning, and a causal AudioVAE/BigVGAN path produces mono 48 kHz waveform output.

mlx-base is not an all-BF16 checkpoint. Its precision policy is:

mlx-int8 applies affine 8-bit quantization with group size 64 only to eligible native qwen.model.* Linear and Embedding modules. Packed weights use U32; their scales and biases remain BF16. Every non-selected path keeps its exact mlx-base dtype. The artifacts are therefore Qwen-selective int8, not whole-model 8-bit conversions.

Usage

Install mlx-speech>=0.5.0 on an Apple Silicon Mac, then load an alias:

pip install "mlx-speech>=0.5.0"

from mlx_speech import tts
from mlx_speech.audio import write_wav

model = tts.load("dots-tts-soar")
result = model.generate(
    "Today the weather is bright and peaceful.",
    reference_audio="reference.wav",
    reference_text="My name is Samantha. I speak clearly and calmly.",
    language="en",
    max_audio_patches=128,
    seed=42,
)
write_wav("output.wav", result.waveform, sample_rate=result.sample_rate)

Waveform streaming

generate_stream() performs bounded-memory waveform streaming and yields mono 48 kHz chunks while the request is running:

for chunk in model.generate_stream(
    "Today the weather is bright and peaceful.",
    reference_audio="reference.wav",
    language="en",
    stream_chunk_patches=4,
):
    consume(chunk.waveform, sample_rate=chunk.sample_rate)

The installed CLI can write the chunks incrementally to one WAV without retaining the complete waveform:

mlx-speech tts \
  --model dots-tts-soar \
  --text "Today the weather is bright and peaceful." \
  --stream \
  --output streamed.wav

The default decoder cadence is one patch, one patch, then four-patch groups. This streams waveform decoding; it does not accept text incrementally or claim universal real-time generation.

Passing reference audio and its matching transcript enables continuation cloning. To use only the CAM++ speaker embedding, omit reference_text:

model = tts.load("dots-tts-mf")
result = model.generate(
    "今天的天气晴朗而平静。",
    reference_audio="reference.wav",
    language="zh",
)

Important generation controls are max_audio_patches, solver_steps, guidance_scale (SOAR only), speaker_scale, language, seed, and eos_threshold. Runtime and conversion code are maintained in the live `mlx-speech` source repository; the sections below record the released artifacts' measured behavior and limitations.

Locally reproduced release gate

App Automaton measured all four artifacts with one fixed macOS English voice and one fixed macOS Mandarin voice. Each voice was tested in continuation and speaker-only modes with a 128-patch cap, seed 42, and EOS threshold 0.8. Qwen3 ASR measured content error rate and CAM++ measured speaker cosine.

The pass thresholds were absolute WER regression ≤ 0.01 and speaker-cosine regression ≤ 0.02. Both int8 artifacts passed. Mandarin error rate used Unicode Han characters as tokens; English used normalized words.

These results are reproduced MLX release measurements, not upstream benchmark claims. The corpus is deliberately small and synthetic, so the results do not establish equal quality across other voices, languages, prompts, seeds, or machines. Full prompts, per-case transcripts, hashes, and methodology are in the checked `2026-07-30 quantization report`.

Provenance

The original source identifiers retained in artifact metadata are rednote-hilab/dots.tts-soar and rednote-hilab/dots.tts-mf; Hugging Face resolves them to the dots-studio repositories above. Original upstream checkpoints are not included in this MLX repository.

Limitations

• —The mlx-speech dots.tts runtime is inference-only; it supports batch output and bounded waveform streaming but does not provide a training path.
• —Continuous autoregressive history grows with the reference and generated sequence. Peak memory can exceed the measurements above for longer inputs or larger patch budgets.
• —No-reference generation follows the target-only schedule, but its random voice was not a quality-supported release-gate mode.
• —English and Mandarin passed the local gate. This card does not publish MLX quality measurements for other languages.
• —Voice identity and pronunciation depend on reference quality, transcript accuracy, text, seed, and generation settings.
• —Quantization passed the fixed release corpus but is not claimed to be lossless or numerically identical to mlx-base.
• —Upstream CUDA real-time measurements do not apply to this MLX implementation. No universal real-time performance claim is made.

Responsible use

High-fidelity voice cloning can enable impersonation and deceptive synthetic speech. Use a voice only with the speaker's authorization. Treat reference recordings as biometric data, restrict their storage and access, and disclose AI-generated audio clearly. Do not use these artifacts for fraud, misinformation, harassment, deceptive attribution, or evasion of consent. Deployers are responsible for applicable law, abuse monitoring, and safeguards appropriate to their users and jurisdiction.

Links

• —Source code: `appautomaton/mlx-speech`
• —Project page: appautomaton.com/mlx-speech

License and attribution

The official dots.tts code and released checkpoints are provided under the Apache License 2.0. These MLX conversions preserve that attribution and are distributed under apache-2.0; consult the upstream release and included metadata when redistributing them.

• —Official source: studio-dots-ai/dots.tts
• —SOAR source weights: dots-studio/dots.tts-soar
• —MeanFlow source weights: dots-studio/dots.tts-mf
• —MLX runtime and conversion code: appautomaton/mlx-speech