FluidInference/neutts-2e-coreml

NeuTTS-2E CoreML

CoreML conversion of neuphonic/neutts-2e (emotional English TTS: Qwen3 236M backbone + NeuCodec decoder) for on-device inference on Apple platforms. Converted by FluidInference; conversion sources live in the mobius repo under models/tts/neutts-2e/coreml/.

Six emotions plus neutral (angry, disgusted, fearful, happy, sad, surprised, neutral) across four fixed speakers (emily, paul, sophie, steven), 24 kHz output.

Files

File	Role	Target
LM-Prefill-T768-M2048-fp16.mlpackage	prompt → last-position logits + KV cache	macOS 14+ / iOS 17+
LM-Decode-M2048-fp16.mlpackage	per-token decode, pass-through KV	macOS 14+ / iOS 17+
LM-Decode-M2048-fp16-stateful.mlpackage	per-token decode, MLState KV	macOS 15+ / iOS 18+
LM-Prefill-T768-M1024-fp16.mlpackage	faster pair, 1024-token cap	macOS 14+ / iOS 17+
LM-Decode-M1024-fp16-stateful.mlpackage	faster pair, 1024-token cap	macOS 15+ / iOS 18+
NeuCodec-Decoder-fp16.mlpackage	speech codes → 24 kHz audio (flexible length 2–2000)	macOS 14+ / iOS 17+
samples/*.pt, samples/*.txt	pre-encoded speaker reference codes + transcripts	—

The M=1024 pair decodes ~23 % faster (7.0 ms/token vs 9.1 on M5 Pro) but caps prompt+generation at 1024 tokens (≈11 s of audio after the emily prompt); use the M=2048 pair for longer utterances. Tokenizer comes from the upstream repo.

Pipeline

text → tokenizer → [prefill] → logits + KV
                        ↓ top-k sampling loop (temp 1.0, k 50), 50 codes/s
                   [decode]  → <|speech_N|> tokens until <|SPEECH_GENERATION_END|>
                        ↓
                   [NeuCodec-Decoder] → 24 kHz waveform

Prompt layout (BPE, no phonemizer): <|TEXT_PROMPT_START|>{ref_text}[<|EMOTION|>]{text}<|TEXT_PROMPT_END|><|SPEECH_GENERATION_START|>{ref codes}. Compute-unit guidance: run the LM on GPU (.all / .cpuAndGPU — the ANE rejects the decode graph), the codec on .cpuAndNeuralEngine (~2× faster than GPU). For streaming, decode the codec in 82-frame windows with 25-frame stride and linear overlap-add (upstream's scheme) — ~550 ms to first audio.

Note: upstream applies a Perth watermark to generated audio in the host app; that postprocessing is not part of these models — hosts should apply it themselves.

Parity & performance (M5 Pro)

• CoreML fp16 LM vs PyTorch fp32: argmax match; teacher-forced replay of a 301-token reference keeps 98.7 % of tokens inside the top-50 sampling support; codec SNR 40–47 dB vs PyTorch across lengths; end audio 41.5 dB vs the PyTorch reference waveform.
• Decode 7.0–9.1 ms/token (109–143 tok/s vs 50 needed for real-time), prefill 33–40 ms warm, codec 12.7–27× RT on ANE.
• Batch ≈ 2× real-time; streaming ≈ 1.6× RT with ~550 ms time-to-first-audio, 0 % ASR round-trip WER on medium/long texts (parakeet-tdt-v3).

License

Inherits the upstream NeuTTS Open License v1.0 (free research use and limited commercial use; paid license required for large-revenue commercial deployments — see LICENSE for exact terms). NeuCodec components follow the neuphonic/neucodec terms.