cguldogan/drama-3-mlx-q8

Moved to mlx-community/fish-audio-drama-3-8bit. The files here are identical and still load, but new updates go there.

Drama 3 MLX 8-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: 8-bit, dequantized from the block-FP8 release
• everything else (attention, shared experts, embeddings, audio tower): bf16, as in the reference
• 42 GB on disk, about 49 GB peak memory while generating, so it needs a 64 GB Mac

It sits between mlx-community/fish-audio-drama-3-4bit (22 GB) and mlx-community/fish-audio-drama-3-bf16 (67 GB): only the routed experts are quantized, at 8 bits, so it stays closest to the reference of the quantized options.

Usage

Drama 3 (fish_drama3) is merged into mlx-audio but not yet in a release; until the next release, install from main:

pip install "mlx-audio @ git+https://github.com/Blaizzy/mlx-audio"

from mlx_audio.tts.utils import load_model

model = load_model("cguldogan/drama-3-mlx-q8")

for result in model.generate(
    text="<|speaker:0|> Did you hear the news? <|speaker:1|> [sighs] That is a real shame.",
    ref_audio="reference.wav",
    ref_text="Transcript of the reference clip.",
    temperature=1.0,
    top_p=0.9,
):
    print(result.audio.shape)  # 44.1 kHz

python -m mlx_audio.tts.generate --model cguldogan/drama-3-mlx-q8 \
    --text "[sighs] Fine. If we miss it, it's on you." \
    --ref_audio reference.wav --ref_text "Transcript of the reference clip." \
    --temperature 1.0 --top_p 0.9 --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 M4 Pro (64 GB), voice-cloned from a 24 kHz English reference, checked with Whisper large-v3-turbo:

• a 4-line, 2-speaker dialog: 12/13 seeds correct at temperature 1.0 / top_p 0.9 (the Fish Audio server's defaults) and 10/14 at mlx-audio's default 0.7 / 0.7; the failures skip lines or ramble. That gap is within noise at this sample size, so either setting is fine
• a cue-heavy line ([sighs] ... [laughs] ...): 2/3 seeds correct; the failed seed stops after the first sentence. Retry with another seed if a take goes silent
• an English single sentence and a Turkish sentence read correctly
• 42.5 GB after load, 49.1 GB peak while generating a 17.5 s dialog

Conversion

Made with mlx-audio's converter:

python -m mlx_audio.tts.models.fish_drama3.convert --src drama-3 --dst drama-3-mlx-8bit \
    --expert-bits 8

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.