Add neutts_nano_tts

[jathinsn27]

Summary

This PR for iOS integration of the NeuTTS Nano text-to-speech model, enabling on-device speech synthesis with voice cloning capabilities. The implementation includes a three-stage pipeline (Backbone → Encoder → Decoder), custom tokenization matching Hugging Face's ByteLevel BPE tokenizer, and integration of espeak-ng for phonemization. Cross-compiling native libraries for iOS and aligning model inputs/outputs with the 3 models from ZeticMLange SDK.

---

1. NeuTTS Nano Architecture Overview

Text Input → Phonemization → Tokenization → Backbone Model → Speech Codes → Decoder → Audio Output
                ↓
         Reference Audio → Encoder → Reference Codes

Components:

1. Backbone Model (neutts_nano): A language model that generates discrete speech tokens from phonemized text
2. Encoder Model (neucodec-encoder): Converts reference audio to discrete codes for voice cloning
3. Decoder Model (neucodec-decoder): Converts discrete codes back to raw audio waveforms

2. Model Input/Output Specifications

2.1 Backbone Model (neutts_nano)

Input:

• input_ids: [1, 128] shape, int32 dtype
  • Tokenized and phonemized text prompt
  • Includes special tokens: <|start_header_id|>, <|end_header_id|>, <|speech_###|>
  • Format: "<|start_header_id|>user<|end_header_id|>\n\n{phonemes}<|speech_###|>"
• attention_mask: [1, 128] shape, int32 dtype
  • Binary mask indicating valid tokens (1) vs padding (0)

Output:

• Logits: [1, 128, vocab_size] shape, float32 dtype
  • Language model logits over vocabulary
  • Contains special speech tokens <|speech_###|> where ### represents discrete audio codes

2.2 Encoder Model (neucodec-encoder)

Input:

• audio: [1, 1, 16000] shape, float32 dtype
  • Mono audio waveform at 16 kHz sample rate
  • Must be exactly 16000 samples (1 second of audio)
  • Values normalized to [-1.0, 1.0] range

Output:

• codes: [1, 1, 50] shape, int32 dtype
  • Discrete audio codes representing the reference audio
  • Used for voice cloning (prosody and timbre transfer)

2.3 Decoder Model (neucodec-decoder)

Input:

• codes: [1, 1, 50] shape, int64 dtype
  • Discrete audio codes from backbone output
  • Must be padded/truncated to exactly 50 codes

Output:

• audio: [1, 1, 24000] shape, float32 dtype
  • Raw PCM audio waveform at 24 kHz sample rate
  • Values in range [-1.0, 1.0]

---

4. Phonemization with Espeak-ng

Initial Approach:

• Attempted to use Hugging Face's transformers library tokenizer directly
• Considered using ZeticMLange to deploy tokenizer as a model (not suitable for text processing)

Phonemization Process:

let phonemes = EspeakPhonemizer.shared.phonemize("Hello world")
// Output: "həlˈoʊ wˈɜːld"

Similar implementation of how its done by neutts: https://github.com/neuphonic/neutts/blob/main/neutts/neutts.py#L11

The phonemes are then tokenized and fed into the backbone model.

Build Settings:

• Add libespeak-ng.a to "Link Binary With Libraries"
• Add -DESPEAK_AVAILABLE to "Other Swift Flags"
• Set header search path to espeak-ng source directory