aoiandroid/neutts-nano-spanish-coreml

NeuTTS Nano Spanish — Core ML

Core ML conversion of neuphonic/neutts-nano-spanish for on-device Spanish TTS on Apple platforms.

Contents

File	Description	Size
NeuTTSNanoSpanish_KVCache_ctx2048.mlpackage	Backbone FP16 (stateful KV Cache)	~1.1 GB
NeuTTSNanoSpanish_KVCache_ctx2048_INT8.mlpackage	Backbone INT8 (weight-quantized)	~600 MB
NeuCodecDecoder.mlpackage	Audio decoder (STFT ri output; Swift does ISTFT)	~100 MB
tokenizer/	Tokenizer files	—
coreml_meta.json	Conversion metadata	—

Requirements

• iOS 18+ / macOS 15+ (backbone requires stateful KV Cache)
• iOS 16+ / macOS 13+ (codec decoder only)

Inference pipeline

1. Phonemize Spanish text with espeak-ng (language: es)
2. Tokenize phonemes + reference speech codes → input_ids
3. Run backbone (prefill + autoregressive decode) → speech token IDs
4. Run NeuCodec decoder → STFT ri spectrum (1, 2*F, N)
5. Swift runtime performs ISTFT → 24 kHz PCM audio

Architecture

• Backbone: LlamaForCausalLM, vocab_size=194256, hidden_size=576, 24 layers, GQA (9 heads / 3 KV heads)
• KV Cache: stateful, max_context_len=2048, size=37.7 MB FP16
• Codec: NeuCodec FSQ, 50 tokens/s, outputs STFT ri components for Swift ISTFT

Notes on NeuCodec decoder

The CoreML NeuCodec decoder outputs STFT real/imaginary (ri) spectrum (1, 2*F, N) rather than raw audio. The Inverse STFT is performed by the Swift runtime using precomputed IDFT matrix and Overlap-and-Add, avoiding CoreML Tools limitations with complex number operations.

Original model

See neuphonic/neutts-nano-spanish for full details, license, and usage instructions.
NeuTTS Nano Spanish is released under the NeuTTS Open License 1.0.