stickerdaniel/drama-3-mlx-4bit

Drama 3 MLX 4-bit

Built with Fish Audio.

Fish Audio Drama 3 converted for mlx-audio on Apple Silicon. This is an unofficial community conversion; Fish Audio did not release or endorse it.

• routed experts: 4-bit, dequantized from the block-FP8 release
• dense projections and audio tower: 8-bit
• 22 GB on disk, about 26 GB peak memory while generating

Usage

Drama 3 support is in review in mlx-audio (#1018). Until it is merged, install from the PR branch:

pip install "mlx-audio @ git+https://github.com/stickerdaniel/mlx-audio@fish-drama3"

from mlx_audio.tts.utils import load_model

model = load_model("stickerdaniel/drama-3-mlx-4bit")

for result in model.generate(
    text="[sighs] Fine. If we miss it, it's on you.",
    ref_audio="reference.wav",
    ref_text="Transcript of the reference clip.",
    stream=True,
    first_chunk_frames=4,
):
    print(result.audio.shape)  # 44.1 kHz chunks

python -m mlx_audio.tts.generate --model stickerdaniel/drama-3-mlx-4bit \
    --text "[sighs] Fine. If we miss it, it's on you." \
    --ref_audio reference.wav --ref_text "Transcript of the reference clip." \
    --temperature 0.7 --top_p 0.7 --top_k 30

Speaker tags (<|speaker:0|> to <|speaker:15|>), multiple references and streaming are described in the model docs.

Quality and speed

On an M3 Max:

• 22 ms per frame (46 ms of audio), about 2x real time
• first streamed audio in 160 to 200 ms once a voice is cached
• teacher-forced against the PyTorch reference: backbone hidden cosine 0.995, semantic top-1 86%

Conversion

Made with mlx-audio's converter at its defaults:

python -m mlx_audio.tts.models.fish_drama3.convert --src drama-3 --dst drama-3-mlx-4bit

NOTICE lists every change to the original weights.

License

The Fish Audio Community License applies, including its use restrictions and its terms for commercial use and hosted services.