T01
dots.tts SOAR
Continuous autoregressive TTS with a 10-step flow-matching solver and classifier-free guidance.
◆ Pure MLX · Apple Silicon · No PyTorch
Speech, rendered on the metal.
An open-source, MLX-native speech library for Apple Silicon. Local text-to-speech, voice cloning, dialogue, sound effects, and automatic speech recognition (ASR), with MLX inference on the Apple GPU. Weights download on first use. Load any model by a short alias.
$ pip install mlx-speech
View on GitHub
15
Speech models
2
Tasks · TTS & ASR
48kHz
Stereo · DramaBox
0
PyTorch at runtime
01 Why pure MLX
The laptop is the whole runtime.
Model inference runs locally through MLX on the Apple GPU, with audio preparation on the CPU where needed. TTS produces playable waveforms. ASR returns text. No PyTorch or cloud inference service is required.
Pure MLX runtime
No torch-backed inference under an MLX label. Weights ship as .safetensors with explicit remapping. Model computation uses the Apple GPU through MLX.
Local & private
Converted weights download once, then run fully offline. Aliases and local checkpoint paths are interchangeable.
One clean interface
tts.load() / asr.load() in Python, or the mlx-speech CLI. Per-family scripts expose the full surface.
02 The catalog
Fifteen models. One loader.
Synthesis, cloning, dialogue, editing, sound effects, and recognition. Every catalog entry links to its converted weights and model-family guide.
Text-to-speech11 modules
int8 · base
T02
dots.tts MeanFlow
Continuous autoregressive TTS with a four-step distilled acoustic solver.
T03
Fish S2 Pro
Dual-AR TTS with voice cloning and inline emotion tags like [excited].
T04
VibeVoice Large
Hybrid LLM-plus-diffusion TTS with voice cloning and long-form delivery.
T05
LongCat AudioDiT
Flow-matching diffusion TTS built on a 3.5B audio diffusion transformer.
T06
OpenMOSS TTS Local
Local-attention, multi-VQ TTS with global + local KV cache.
T07
MOSS-TTSD
Delay-pattern dialogue TTS for multi-speaker conversation.
T08
OpenMOSS Sound Effect
Text-to-sound-effect generation. Turn "rolling thunder with rainfall" into a clip.
T09
Step-Audio-EditX
Voice cloning plus emotion and style edits on existing speech.
T10
DramaBox
Resemble flow-matching diffusion TTS at 48 kHz stereo.
T11
FireRedTTS3 Base
Multilingual zero-shot voice cloning with mono 24 kHz output. Complete BF16 inference bundle. Requires the current GitHub install.
Speech-to-text04 modules
A01
Cohere Transcribe
Multilingual ASR with strong long-audio behavior.
A02
Qwen3-ASR-1.7B
English, Chinese, and mixed Chinese/English recognition.
A03
Nemotron 3.5 ASR Streaming
Cache-aware multilingual streaming ASR with bounded encoder and decoder state.
A04
IBM Granite Speech 4.0 1B
Selective-int8 Granite LM with the acoustic encoder and QFormer retained in BF16.
Published weights live under appautomaton on Hugging Face and load by alias or full repo id, with an artifact subdirectory where required. Original checkpoints can also be loaded from local paths when a family guide documents that path.
03 Quickstart
Install. Load. Generate.
Requires an Apple Silicon Mac (M1 or later) and Python 3.13+. Weights download on first use. FireRedTTS3 Base currently requires the GitHub installation.
import mlx_speech
# Text-to-speech
model = mlx_speech.tts.load("fish-s2-pro")
result = model.generate("Hello from mlx-speech!")
# result.waveform: mx.array · result.sample_rate: int
# Voice cloning with emotion tags
result = model.generate(
"[excited] This is amazing!",
reference_audio="reference.wav",
reference_text="Transcript of the reference.",
)
# Speech-to-text
asr = mlx_speech.asr.load("qwen3-asr-1.7b")
print(asr.generate("audio.wav").text)
$ pip install mlx-speech
$ mlx-speech tts --model fish-s2-pro --text "Hello!" -o out.wav
✓ out.wav
$ mlx-speech asr --model qwen3-asr-1.7b --audio speech.wav
$ mlx-speech tts --list-models