Split ONNX graphs (LM + audio decoder) exported with practical fixed shapes (see onnx/onnx_shapes.json).
The original SentencePiece tokenizer (tokenizer.model) plus a 1:1 verified fast tokenizer.json for swift-transformers.
The upstream config.json, exposing all token IDs and model-side constants used by the prompt builder.

Repo layout

.
├── onnx/
│   ├── moss_tts_nano_lm_fp32.onnx        # ~490 MiB — use on macOS (CoreML EP) / Mac-class iPad
│   ├── moss_tts_nano_lm_fp16.onnx        # ~250 MiB — default for iPhone
│   ├── moss_tts_nano_decode_fp32.onnx    # ~62 MiB
│   ├── moss_tts_nano_decode_fp16.onnx    # ~32 MiB
│   └── onnx_shapes.json                  # prompt_len / frames / sample rate / special tokens
├── tokenizer/
│   ├── tokenizer.json                    # fast tokenizer (BPE + byte fallback) verified 1:1 against SPM
│   ├── tokenizer.model                   # original SentencePiece model
│   ├── tokenizer_config.json
│   └── special_tokens_map.json
└── config/
    └── config.json                       # upstream MOSS-TTS-Nano config

Fixed shapes (baked into the ONNX graphs)

field	value	meaning
`prompt_len`	512	Max `prompt_input_ids` rows (text tokens + MOSS template overhead). Pad/truncate per chunk.
`n_vq`	16	Number of audio codebooks.
`n_vq_plus_one`	17	Columns in `prompt_input_ids` (1 text column + 16 audio codebook columns, all set to `audio_pad_token_id` at the text-only prompt stage).
`max_new_frames`	32	Frames the LM can emit per invocation.
`max_audio_samples`	1,920,000 (≈ 40 s @ 48 kHz)	Upper bound on the decoder output length.
`sample_rate`	48,000 Hz	Output WAV sample rate.
`channels`	2	Stereo output.

Long inputs must be split into prompt-budget-sized chunks in the client (sentence-aware, with a single MOSS prompt overhead per chunk) and the resulting WAV segments concatenated. The on-device iOS package does this via MossTTSNanoTextChunker.

Runtime contract (on-device iOS / macOS client)

Build prompt_input_ids / attention_mask on-device using the fast tokenizer.json (no server roundtrip).
- See MossTTSNanoPromptBuilder — Swift port of MossTTSNanoForCausalLM.build_inference_input_ids(mode="continuation").
- Pads / truncates to prompt_len = 512 and fills audio codebook columns with audio_pad_token_id.
Run moss_tts_nano_lm_*.onnx to produce audio_codes of shape [1, max_new_frames, n_vq].
Run moss_tts_nano_decode_*.onnx on a sliding window of frames to obtain raw 48 kHz stereo PCM.
Concatenate segment WAVs for multi-chunk inputs.

Export provenance

Exported via Notebooks/MOSS_TTS_Nano/export_moss_tts_nano_e2e_onnx.py (torchscript backend, opset 17) with:

python export_moss_tts_nano_e2e_onnx.py \
    --mode split \
    --prompt-len 512 \
    --max-new-frames 32 \
    --max-audio-samples 1920000 \
    --write-shapes-json

FP16 graphs produced with onnxconverter_common.float16.convert_float_to_float16(keep_io_types=False).

Token IDs in onnx_shapes.json reflect the upstream config at export time (OpenMOSS-Team/MOSS-TTS-Nano-100M commit 44502f80).

License

Apache-2.0, same as the upstream MOSS-TTS-Nano-100M and MOSS-Audio-Tokenizer-Nano releases.

Acknowledgements

Thanks to the OpenMOSS team for releasing the MOSS-TTS-Nano and MOSS-Audio-Tokenizer-Nano weights under a permissive license that makes on-device deployment feasible.

aoiandroid/moss-tts-nano-onnx-fp32

on-device iOS client · MOSS-TTS-Nano (ONNX for iOS / macOS)