MLX-TADA

TADA speech synthesis on Apple Silicon via MLX.

Also available on PyPI: pip install mlx-tada

Setup

cd apple
uv venv
uv pip install -e .

For auto-transcription of reference audio (optional):

uv pip install mlx-whisper

Weights

Download a reference audio clip:

curl -O "https://storage.googleapis.com/hume_reference_speakers/ljspeech.wav"

Pre-converted weights are downloaded and cached automatically. You still need gated access to Llama 3.2 for the tokenizer:

from mlx_tada import TadaForCausalLM, save_wav

model = TadaForCausalLM.from_pretrained("HumeAI/mlx-tada-3b", quantize=4)
ref = model.load_reference("ljspeech.wav")
out = model.generate("Hello, this is a test of TADA speech synthesis.", ref)
save_wav(out.audio, "output.wav")

Available models:

• HumeAI/mlx-tada-1b — 1B English-only (~4.3 GB)
• HumeAI/mlx-tada-3b — 3B multilingual (~8.9 GB)

Offline Use

To download weights locally for offline inference:

from huggingface_hub import snapshot_download
snapshot_download("HumeAI/mlx-tada-3b", local_dir="./weights/3b")

Then load from the local path:

model = TadaForCausalLM.from_weights("./weights/3b", quantize=4)

Generate Speech

CLI

uv run python -m mlx_tada.generate \
  --weights ./weights/3b \
  --audio ljspeech.wav \
  --text "The history of artificial intelligence is a fascinating journey that spans decades of research and innovation. It all began in the 1950s when pioneers like Alan Turing first posed the question of whether machines could think." \
  --output output.wav

With 4-bit quantization (10x faster, 60% less memory):

uv run python -m mlx_tada.generate \
  --weights ./weights/3b \
  --audio ljspeech.wav \
  --text "The history of artificial intelligence is a fascinating journey that spans decades of research and innovation. It all began in the 1950s when pioneers like Alan Turing first posed the question of whether machines could think." \
  --quantize 4 \
  --output output.wav

Python

from mlx_tada import TadaForCausalLM, save_wav

model = TadaForCausalLM.from_pretrained("HumeAI/mlx-tada-3b", quantize=4)
ref = model.load_reference("ljspeech.wav")
out = model.generate("The history of artificial intelligence is a fascinating journey that spans decades of research and innovation. It all began in the 1950s when pioneers like Alan Turing first posed the question of whether machines could think.", ref)
save_wav(out.audio, "output.wav")

# out.audio     - numpy float32 array (24kHz)
# out.duration  - audio duration in seconds
# out.rtf       - real-time factor
# out.num_tokens

Inference Options

Control generation behavior with InferenceOptions:

from mlx_tada import TadaForCausalLM, InferenceOptions, save_wav

model = TadaForCausalLM.from_weights("./weights/3b", quantize=4)
ref = model.load_reference("ljspeech.wav")

opts = InferenceOptions(
    acoustic_cfg_scale=1.6,
    duration_cfg_scale=1.0,
    num_flow_matching_steps=10,
    time_schedule="logsnr",
    cfg_schedule="cosine",
)

out = model.generate(text="Hello world, today is a nice day.", reference=ref, inference_options=opts)
save_wav(out.audio, "output.wav")

The following inference options from the PyTorch version are not currently supported in MLX:

• speed_up_factor
• num_acoustic_candidates
• scorer
• negative_condition_source
• text_only_logit_scale
• spkr_verification_weight

Speech Continuation

Use num_extra_steps to let the model generate speech beyond the provided text. The model continues speaking naturally and stops when it produces an end-of-sequence token:

from mlx_tada import TadaForCausalLM, InferenceOptions, save_wav

model = TadaForCausalLM.from_weights("./weights/3b", quantize=4)
ref = model.load_reference("ljspeech.wav")

opts = InferenceOptions(
    acoustic_cfg_scale=1.6,
    num_flow_matching_steps=10,
    time_schedule="logsnr",
)

out = model.generate(
    text="The history of artificial intelligence is a fascinating journey that spans decades of research and innovation.",
    reference=ref,
    inference_options=opts,
    num_extra_steps=50,
)
save_wav(out.audio, "output.wav")

Save and Reuse References

from mlx_tada import Reference

ref = model.load_reference("ljspeech.wav")
ref.save("speaker.npz")

ref = Reference.load("speaker.npz")
out = model.generate("Reusing the same voice.", ref)

Save Audio

from mlx_tada import save_wav
save_wav(out.audio, "output.wav")

Debug Logging

DEBUG=1 uv run python -m mlx_tada.generate \
  --weights ./weights/3b \
  --audio ljspeech.wav \
  --text "Hello"

from mlx_tada import setup_logging

setup_logging()

Running Tests

MLX_WEIGHTS=./weights/1b uv run pytest tests/ -v