Skip to content

VoxCPM2

VoxCPM2 is OpenBMB's 2B-parameter, multilingual tokenizer-free text-to-speech model. The MLX implementation produces 48 kHz audio and supports standard synthesis, voice design from a text description, voice cloning from reference audio, and speech continuation.

Model Variants

Model Format Size HuggingFace
mlx-community/VoxCPM2-bf16 bfloat16 4.96 GB Model Card
mlx-community/VoxCPM2-8bit 8-bit 3.23 GB Model Card
mlx-community/VoxCPM2-4bit 4-bit 2.30 GB Model Card

Quick Start

mlx_audio.tts.generate \
    --model mlx-community/VoxCPM2-8bit \
    --text "Hello, this is VoxCPM2 running on Apple Silicon." \
    --output_path outputs

from mlx_audio.tts import load

model = load("mlx-community/VoxCPM2-8bit")

result = next(
    model.generate("Hello, this is VoxCPM2 running on Apple Silicon.")
)
audio = result.audio  # 48 kHz mono waveform as an MLX array

Voice Design

Describe the desired speaker and delivery with instruct; no reference audio is required.

mlx_audio.tts.generate \
    --model mlx-community/VoxCPM2-8bit \
    --text "Welcome. I hope you enjoy the presentation." \
    --instruct "A warm, confident young woman speaking at a relaxed pace" \
    --output_path outputs

result = next(
    model.generate(
        text="Welcome. I hope you enjoy the presentation.",
        instruct="A warm, confident young woman speaking at a relaxed pace",
    )
)

Voice Cloning

Pass a clean reference recording to synthesize the target text in the reference speaker's voice.

mlx_audio.tts.generate \
    --model mlx-community/VoxCPM2-8bit \
    --text "This sentence uses the voice from the reference recording." \
    --ref_audio speaker.wav \
    --ref_text "The exact transcript of speaker.wav." \
    --output_path outputs

result = next(
    model.generate(
        text="This sentence uses the voice from the reference recording.",
        ref_audio="speaker.wav",
    )
)

ref_text is part of the common MLX-Audio CLI voice-cloning interface. Supplying it prevents the CLI from running automatic reference-audio transcription; VoxCPM2 conditions directly on the reference audio.

Speech Continuation

For long-form generation, provide an earlier audio clip and its transcript. The new text should begin where prompt_text ends.

result = next(
    model.generate(
        text=" The story continues from here.",
        prompt_text="This is the opening of the story.",
        prompt_audio="opening.wav",
    )
)

Continuation is currently available through the Python API. Reference voice cloning can be combined with continuation by also passing ref_audio.

Generation Settings

Parameter Default Description
inference_timesteps 10 Diffusion steps; higher values trade speed for quality
cfg_value 2.0 Classifier-free guidance strength
instruct None Natural-language voice description for voice design
ref_audio None Reference audio path or waveform for voice cloning
prompt_text None Transcript of prompt_audio for continuation
prompt_audio None Previous audio clip used as continuation context
warmup_patches 0 Extra conditioning patches excluded from the output
max_tokens 2000 Maximum number of audio patches to generate

The CLI exposes --ddpm_steps as an alias for inference_timesteps and --cfg_scale as an alias for cfg_value. VoxCPM2 enforces a minimum guidance value of 2.0.

License

VoxCPM2 is released under the Apache License 2.0.