StillDeadcode/radiance-breeze-tts2-bf16

Breeze-TTS-2, bf16 — radiance container

Derived from Breeze TTS 2 by BreezeBlue and licensed for research and non-commercial use only.

BreezeBlue/Breeze-TTS-2 — text to speech, in a voice described in words — as a single .rad container for the radiance inference engine (AMD RDNA4, ROCm).

File	breeze-tts2-bf16.rad — 5.53 GiB
Weights	the backbone, the depth decoder and the codec's linears and convolutions bf16; the codec's norms, layer scales and snake parameters f32
Audio	24 kHz mono, 16-bit
Needs	radiance 1.8.0 or newer (Breeze-TTS-2 support comes with 1.8.0); one card

Serve

radiance --model breeze-tts2-bf16.rad --host 0.0.0.0 --port 8000

curl -s http://localhost:8000/v1/audio/speech -H 'Content-Type: application/json' -d '{
  "input": "It was a bright cold day in April.",
  "instructions": "A calm, warm female narrator, speaking slowly.", "guidance_scale": 4}' -o april.wav

The server speaks OpenAI's POST /v1/audio/speech and vLLM-Omni's realtime speech WebSocket:

• input is the text to speak. instructions, optional, describe the voice; BreezeBlue suggests guidance_scale 4 with them.
• duration (seconds) or max_new_tokens (frames, 12.5 a second) bounds the length, up to two minutes. seed makes a speech repeatable; response_format is wav or pcm.
• "stream_format": "audio" sends the audio as it is made; "stream_format": "sse" (or "stream": true) sends it as speech.audio.delta events.
• /v1/audio/speech/stream (WebSocket) speaks text while it is still being written, each sentence as soon as it is complete.
• Cloning a voice from a recording is not served.

On one Radeon AI PRO R9700 (gfx1201): 5.1 s of speech in 3.1 s; a stream's first audio 0.21 s after the request; four requests at once, 43 s of speech in 12.1 s; over the WebSocket, a sentence's first audio 0.20 s after its last word.

Fed the same codes as BreezeBlue's reference pipeline, the backbone's hidden state is 1.0% from the reference's (the reference's own bf16 against its f32: 0.9%), and the codec's audio is 38.4 dB SNR from the reference's f32 decode.

How this file was made

rad-convert BreezeBlue/Breeze-TTS-2 --recipe breeze-tts2-bf16.recipe -o breeze-tts2-bf16.rad

BreezeBlue/Breeze-TTS-2 is the downloaded repository; the codec in its audio_tokenizer/ is read with it. The recipe (breeze-tts2-bf16.recipe in this repository) names what changes dtype; everything else is the checkpoint's own bf16:

dep.heads.weight                     cast dtype=bf16
codec.*.n1.weight                    cast dtype=f32
codec.*.n2.weight                    cast dtype=f32
codec.*.ls1.weight                   cast dtype=f32
codec.*.ls2.weight                   cast dtype=f32
codec.norm.weight                    cast dtype=f32
codec.up*.ln.*                       cast dtype=f32
codec.up*.gamma                      cast dtype=f32
codec.*.alpha                        cast dtype=f32
codec.*.beta                         cast dtype=f32
codec.*                              cast dtype=bf16

License

Breeze TTS 2 is licensed under the BreezeBlue Research and Non-Commercial License Agreement (LICENSE in this repository): research and non-commercial use only, and the agreement counts hosting the model as an API, an application or a website as commercial use. NOTICE carries its attribution. This container is a modified copy: the weights are repacked into one .rad file, the depth decoder's output heads are transposed, and the codec's linears and convolutions are cast from f32 to bf16, its codebooks divided by their usage counts first. The audio tokenizer is based on Qwen3-TTS by the Alibaba Qwen team (Apache License 2.0).