FluidInference / inflect-v2-coreml

Inflect v2 CoreML (Micro + Nano)

Fixed-shape CoreML conversion of Inflect-Micro-v2 (9.36M params) and Inflect-Nano-v2 (3.97M params) — ultra-tiny VITS-family end-to-end English TTS, 24 kHz mono, fp16.

Converted by FluidInference (mobius); conversion pipeline, parity checks, and benchmark harness live in the models/tts/inflect-v2/coreml/ directory of the mobius repo.

Contents

micro/            9.4M params, ~19 MB fp16 total per graph set
  encoder.mlpackage             tokens+mask -> m_p, logs_p, logw   [t_text=512]
  synthesizer_f{256,384,512,640,768,896,1024,2048}.mlpackage
                                z_p+mask -> waveform  [bucket frames x 256 samples]
  upstream_config.json
nano/             4.0M params, ~8 MB fp16 per graph set (same layout)
symbols.json      keithito-style symbol table (178 symbols, add_blank interspersal)
LICENSE           Apache-2.0 (upstream)

Inference pipeline

Two deterministic models with everything stochastic or dynamically shaped on the host:

1. espeak-ng en-us phonemization (with stress) → symbol ids → intersperse blanks ( pad_id=0 ) → pad to 512.
2. encoder : → m_p , logs_p , logw .
3. Host: w = ceil(exp(logw))/speed ; repeat-expand m_p / logs_p to y_len frames; z_p = m_p + randn * exp(logs_p) * noise_scale (default 0.667).
4. Pick the smallest synthesizer bucket ≥ y_len ; zero-pad z_p , mask valid frames.
5. synthesizer : reverse coupling flow + HiFiGAN → waveform; trim to y_len * 256 samples.

Buckets are fixed-shape for ANE compatibility; at 128-frame granularity padding overhead is within ~5% of exact-shape synthesis.

Benchmarks (M5 Pro, macOS 26.6, GPU, MiniMax-English 100 phrases)

* Parakeet TDT v3 roundtrip, same scoring path as the FluidAudio TTS benchmarks. fp16 parity vs the PyTorch reference: audio correlation > 0.9999, predicted durations bit-identical.

Notes

• ANE: the synthesizer's waveform-rate tensors exceed the ANE width limit (W ≤ 65536) above the 256-frame bucket; f256 reaches 77% ANE residency, larger buckets run on GPU (which is faster at every size on M-series).
• English only, one voice per variant, no cloning. Sample rate 24 kHz.
• Upstream weights and frontend: Apache-2.0, © the Inflect authors. This repo redistributes converted weights under the same license.