Paradee-8M (beta) — ParadeeManager (FluidInference/paradee-8m-coreml). Sahil Mahendrakar's Paradee distills Kokoro-82M into an 8M-parameter, single-voice (af_heart) American English TTS model. The Core ML build is 12 MB (int8, default) or 34 MB (fp32) and reuses the Kokoro English frontend.
| MiniMax English, 100 phrases, M5 Pro, CPU | Paradee int8 | Paradee fp32 |
|---|---|---|
| Download | 12 MB | 34 MB |
| WER / CER (Parakeet round trip) | 1.20 % / 0.14 % | 1.20 % / 0.14 % |
| RTFx (incl. G2P) | ~100× | ~100× |
| Median synth time per phrase | 77 ms | 77 ms |
On the same inputs, the Core ML models run ~150× real time against ~20× for the upstream ONNX on one thread (~47× on all 15 threads), with identical Whisper transcripts.
let manager = ParadeeManager()
try await manager.initialize()
let samples = try await manager.synthesize(text: "Hello from FluidAudio.") // 24 kHz Float32CLI: tts --backend paradee [--variant int8|fp32] [--speed] [--seed] [--phonemes]; tts-benchmark --backend paradee. Runs on .cpuOnly (default) or .cpuAndNeuralEngine; GPU compute units are rejected (the LSTMs abort in MPSGraph). See Paradee.md.
Also in this release: Kokoro ANE English now applies misaki's ɾ → T flap step, so words like "metal" no longer come out as "mattle" (#999), and updated Kokoro v3 benchmark, compute-routing and ANE-profiling docs (#1000).
What's Changed
- feat(tts): Paradee-8M CoreML backend [beta] by @Alex-Wengg in #998
- fix(tts/kokoro-ane): apply misaki's ɾ → T flap step to English phonemes by @Alex-Wengg in #999
- docs(kokoro): update v3 benchmarks, compute routing and ANE profiling in #1000
Full Changelog: v0.17.6...v0.17.7