Breeze TTS 2

audio_tokenizer/LICENSE · Apache-2.0 · Qwen/Qwen3-TTS-Tokenizer-12Hz

Apple M5 Pro · 48 GiB · TTFA / RTF / RSS

2026-10-30 · en · 8bit=3.20 s · bf16=3.20 s · seed=42 · CFG=4 · firstChunkFrames=3 · chunkFrames=25 · MLX cache=2 GiB · warm=3

MLX	TTFA p50	RTF	RSS (GiB)	Peak physical footprint (GiB)
8bit	178 ms	0.535	3.92	7.23
bf16	272 ms	0.868	6.06	9.22

WAV=45 · distinct=11 · 8bit PCM SHA256=baseline · unit=1933 · skipped=42 · E2E=15 · Qwen3-ASR-0.6B-MLX-4bit

BF16 now measures 0.868 RTF on the short sentence (baseline: 1.258) and 0.853 on the 9.20 s paragraph (baseline: 1.280). One initial frame gives first PCM at 138 ms. All 11 distinct final waveforms have zero normalized WER/CER in independent ASR; 8-bit default-chunk output is byte-identical to the baseline. BF16 rounding can change seeded audio. Set SPEECH_SWIFT_BREEZE_GEMV=0 before loading to compare the original projection path.

For lower playback startup, use one initial frame, two follow-up frames, and a 1 GiB MLX cache budget. The 9.04 s 8-bit paragraph reaches the estimated CLI prebuffer at 190 ms, with no extra buffering required by recorded chunk arrivals. This excludes render callback and output-device latency.

2026-10-30 · 8bit · 9.04 s · firstChunkFrames=1 · chunkFrames=2 · MLX cache=1 GiB · PCM=103 ms · buffer=190 ms · RTF=0.761 · peak=6.10 GiB

speech speak "The quick brown fox jumps over the lazy dog." --engine breeze \
  --instruct "A clear, calm English narrator." --first-chunk-frames 1 \
  --chunk-frames 2 --breeze-cache-limit-mib 1024 --stream -o fast-start.wav

TTFA excludes model loading and speaker playback. Results depend on hardware, text, and settings. The shared model policy caps reusable MLX cache at the smaller of 2 GiB and one quarter of physical memory, honoring tighter application budgets. --breeze-cache-limit-mib can lower this process-wide ceiling; 0 disables reusable caching. Live model tensors are not limited.

English/Chinese TTS with voice design, reference cloning and direction, and streamed audio (MLX BF16/8-bit, 24 kHz; research/non-commercial license).

BreezeTTS · 24,000 Hz · 16 × 2,048 · 80 ms · MLX BF16 / 8-bit

Usage

speech speak "The quick brown fox jumps over the lazy dog." --engine breeze \
  --instruct "A clear, calm English narrator." --stream -o design.wav

speech speak "The weather is beautiful today." --engine breeze \
  --voice-sample reference.wav --breeze-reference-text "Exact transcript." \
  --breeze-seed 42 -o clone.wav

speech speak "We need to discuss what happened." --engine breeze \
  --voice-sample reference.wav --breeze-reference-text "Exact transcript." \
  --instruct "Speak slowly, with a serious tone." --stream -o directed.wav

Swift API

import BreezeTTS

let model = try await BreezeTTSModel.fromPretrained()
let samples = try await model.generate(
    text: "The quick brown fox jumps over the lazy dog.",
    instruction: "A clear, calm English narrator.")

try await model.generateStreaming(text: "Hello, world!", instruction: "A warm voice.") { chunk in
    print(chunk.sampleRate, chunk.samples.count, chunk.isFinal)
}

Options

--breeze-model-id: aufklarer/Breeze-TTS-2-MLX-8bit
--breeze-bundle-dir
--breeze-cache-limit-mib: Optional process-wide reusable MLX cache budget in MiB; 0 disables caching.
--breeze-reference-text
--breeze-temperature: 0.9
--breeze-top-k: 50
--breeze-top-p: 1
--breeze-cfg-scale: 4
--breeze-max-frames: 750
--breeze-seed: 42
--first-chunk-frames: 3
--chunk-frames: 25

License

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

BreezeBlue Research and Non-Commercial License · Breeze TTS 2 MLX 8-bit · Breeze TTS 2 MLX BF16

Source and conversion

speech-swift / BreezeTTS · export_breeze_mlx.py · WER / CER / RSS / TTFA / RTF