AEmotionStudio/moss-tts-models

MOSS-TTS v1.5 family — MAESTRO mirror

Redistribution of the OpenMOSS MOSS-TTS v1.5 checkpoints and their codecs, 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-TTS v1.5 is zero-shot voice-cloning text-to-speech in 31 languages: clone a speaker from a few seconds of reference audio (no transcript) or use the model's own voice, with explicit language tags, inline [pause X.Ys] marks and token-level duration control.

lane	upstream	revision	what
local_v15/	MOSS-TTS-Local-Transformer-v1.5	be7766a6735b98bd793f7c79fb720b4d0f5d13b8	4B MossTTSLocal — Qwen3-4B backbone + a 1-layer local transformer over 12 RVQ codebooks
local_v15/codec/	MOSS-Audio-Tokenizer-v2	f6e20e543b33d2c252a7ef71bdf8aa71e5ff9169	native 48 kHz stereo, 12.5 Hz, 2B params
delay_v15/	MOSS-TTS-v1.5	cdd3b911b1585e3f2dbc7775ef10f9926f58850a	8B MossTTSDelay — Qwen3-8B backbone + delay-pattern heads over 32 RVQ codebooks
delay_v15/codec/	MOSS-Audio-Tokenizer	3cd226ba2947efa357ef453bcad111b6eafba782	24 kHz mono, 12.5 Hz, 1.77B params

Layout

local_v15/model/   config.json, model.safetensors (bf16, 9.10 GB — VERBATIM upstream),
                   tokenizer + config files
local_v15/codec/   config.json (upstream + the shipped codec_weight_dtype)
                   encoder.safetensors   — encoder.* tensors
                   decoder.safetensors   — decoder.* tensors
                   quantizer.safetensors — quantizer.* tensors (always fp32)
                   model.safetensors.index.json — maps every tensor → the three files
                   sha256.json           — upstream shard hashes + per-tensor provenance
delay_v15/model/   config.json, model-0000N-of-00004.safetensors (bf16, 17.0 GB — VERBATIM),
                   model.safetensors.index.json, tokenizer + config files
delay_v15/codec/   config.json, decoder.safetensors (decoder.* + quantizer.*, fp32),
                   encoder.safetensors (encoder.*, fp32), model.safetensors.index.json,
                   sha256.json — byte-identical to upstream (per-tensor sha256 proof)

Why the codecs are re-sharded

Voice cloning encodes the reference clip once and decodes the take once — never both at the same time. The codecs are re-partitioned by top-level module so MAESTRO can materialise one half at a time and move it on and off the GPU around its single call; on a 12 GB card that is the difference between the 4B language model fitting and not. Each lane's codec/sha256.json carries the upstream shard hashes and a per-tensor sha256 so the split is verifiable: the v1 codec and the v2 quantizer are byte-identical to upstream; if local_v15/codec/config.json says codec_weight_dtype: bf16, the v2 encoder/decoder halves were cast fp32 → bf16 (upstream's own default codec_weight_dtype for this model) and both the upstream and the shipped digest are recorded.

What was dropped

The upstream repos' remote-code .py files. MAESTRO vendors the GitHub moss_tts_local_v1.5/ and moss_tts_delay/ graphs and both codec implementations (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 (the model repos carry it):

from transformers import AutoModel, AutoProcessor
processor = AutoProcessor.from_pretrained("OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5",
                                          trust_remote_code=True)
model = AutoModel.from_pretrained("AEmotionStudio/moss-tts-models", subfolder="local_v15/model",
                                  trust_remote_code=True, torch_dtype="bfloat16")

Inside MAESTRO: backend/ai/models/moss_tts.py (one runner, both lanes; codec halves via ai/moss_tts_vendor/audio_tokenizer_v2/loading.py and 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.