aufklarer/Higgs-TTS-3-4B-MLX-bf16

Higgs TTS 3 — MLX bf16

Native MLX bundle of Higgs TTS 3 — a ~4B conversational text-to-speech model built on a Qwen3-4B decoder with a fused 8-codebook audio interface (8 x 1026 vocab at 25 fps, MusicGen-style delay pattern) and an embedded Higgs codec, producing 24 kHz speech. Zero-shot voice cloning from a short reference clip plus transcript, 100+ languages, and inline control tokens for emotion, style, sound effects, and prosody.

Part of soniqo.audio — an on-device speech toolkit for Apple Silicon. Consumed by the open-source speech-swift library (module HiggsTTS).

Bundle size: 9321 MB

Capabilities

• Zero-shot voice cloning — one short reference clip plus its transcript
• 100+ languages (85 at WER/CER < 5 per the upstream card)
• Inline control tokens — <|emotion:...|> (21), <|style:...|> (3), <|sfx:...|> (9), <|prosody:...|> (10), placed directly in the text
• 24 kHz output at 25 fps via the embedded Higgs codec
• Conversational delivery — built for voice chat, speaks rather than reads

Model

Field	Value
Backbone	Qwen3-4B decoder (36 layers, hidden 2560, GQA 32/8)
Audio interface	8 codebooks x 1026 vocab, 25 fps, delay pattern
Codec	embedded (DAC-style acoustic enc/dec + semantic distillation + 8-layer RVQ)
Sample rate	24 kHz
Context	8192 tokens
Precision	bf16, upstream key names preserved

Files

File	Description
model.safetensors	Full system: backbone, fused codebook embedding/head, codec (bf16)
config.json	Upstream architecture config
tokenizer.json, tokenizer_config.json	Qwen tokenizer with Higgs special/control tokens
chat_template.jinja	Upstream chat template
soniqo_manifest.json	Bundle manifest: source, license posture, file hashes

Validation

Apple Silicon roundtrip gate (cloned voice, transcribed with Qwen3-ASR): English 0.0% WER, Mandarin 0.0% CER, Spanish 0.0% WER, Japanese 3.0% CER at RTF 0.8-0.9 in the Python MLX reference implementation.

Source

This bundle preserves the upstream weights at bosonai/higgs-tts-3-4b (formerly bosonai/higgs-audio-v3-tts-4b), built by Boson AI on the Qwen3-4B-Base backbone. Technical blog: https://boson.ai/blog/higgs-audio-v3-tts.

License

Boson Higgs TTS 3 Research and Non-Commercial License (license: other), inherited from upstream. Research and non-commercial use; a Creator Use Grant permits monetized creator content with attribution to "Boson AI's Higgs Audio". Production, hosted APIs, or embedding in products requires a commercial license from Boson AI.

Responsible use

Voice cloning capability is included. Users are responsible for obtaining consent for any voice that is cloned and for not using the model to impersonate individuals without their permission, generate disinformation, or commit fraud.