AEmotionStudio/moss-voicegenerator-models

MOSS-VoiceGenerator — MAESTRO mirror

Redistribution of OpenMOSS-Team/MOSS-VoiceGenerator (revision 97521ec2b6f3ec5026ac1f5751f8fc302d82c2d4) and its codec OpenMOSS-Team/MOSS-Audio-Tokenizer (revision 3cd226ba2947efa357ef453bcad111b6eafba782), packaged for the MAESTRO audio workstation's offline runtime. All credit to the OpenMOSS team (Fudan University) — code and weights are Apache-2.0 (MOSS-TTS, MOSS-Audio-Tokenizer).

MOSS-VoiceGenerator designs a speaker timbre from a free-form text description — no reference audio — and reads text in that voice (English and Chinese) with expressive, emotional delivery. 1.7B MossTTSDelay (Qwen3-1.7B backbone + 16 delay-pattern RVQ heads) over the MOSS-Audio-Tokenizer codec (24 kHz mono, 12.5 Hz frames, 1.77B parameters).

Layout

model/   config.json, model.safetensors (bf16, 4.23 GB — VERBATIM upstream),
         tokenizer.json, tokenizer_config.json, vocab.json, merges.txt,
         added_tokens.json, special_tokens_map.json, chat_template.jinja
codec/   config.json (verbatim upstream)
         decoder.safetensors   — decoder.* + quantizer.* tensors (fp32)
         encoder.safetensors   — encoder.* tensors (fp32)
         model.safetensors.index.json — maps all 1600 tensors → the two files
         sha256.json           — upstream shard hashes + per-tensor sha256 proof

Why the codec is re-sharded

Upstream ships the codec as two 4.99 GB + 2.10 GB shards with encoder and decoder tensors mixed. MAESTRO's voice-design runner never encodes audio, so it only needs the decoder + quantizer half (~3.5 GB). The mirror re-partitions the same checkpoint by tensor so that half can be downloaded on its own. Every tensor is byte-identical to upstream — codec/sha256.json carries a per-tensor sha256 computed from the upstream shards and re-verified against the re-sharded files. Loading decoder.safetensors + encoder.safetensors through the index gives back the full upstream state dict.

What was dropped

The upstream repos' remote-code .py files. MAESTRO vendors the GitHub moss_tts_delay/ graph and the codec implementation (patched for transformers 4.57.x) and never uses trust_remote_code. The upstream READMEs are replaced by this card.

Loading

With the upstream code:

from transformers import AutoModel, AutoProcessor
processor = AutoProcessor.from_pretrained("OpenMOSS-Team/MOSS-VoiceGenerator", trust_remote_code=True,
                                          codec_path="AEmotionStudio/moss-voicegenerator-models/codec")

Inside MAESTRO: backend/ai/models/moss_voicegenerator.py (decoder-only codec load via ai/moss_tts_vendor/audio_tokenizer/loading.py).

License

Apache License 2.0 — © OpenMOSS / Fudan University. This mirror adds no restrictions and claims no ownership. Please cite the upstream projects.