No official torrent yet โ requesting it tells the flock what to host next. You cannot seed a catalog-only page.
Cataloged from OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5
All models by OpenMOSS-Team โย ย ย ย
MOSS-TTS-Local-Transformer-v1.5 is continued from MOSS-TTS-Local-Transformer-v1.0. It preserves the main 1.0 capabilities, including zero-shot voice cloning, long-form speech generation, token-level duration control, Pinyin/IPA pronunciation control, multilingual synthesis, and code-switching. For the full 1.0 feature walkthrough, input schema, and evaluation tables, please refer to the MOSS-TTS-Local-Transformer-v1.0 README.
Compared with MOSS-TTS-Local-Transformer-v1.0, v1.5 focuses on the following improvements:
[channels, samples] tensor returned by processor.decode(...) directly.language field is omitted, v1.5 may improve some languages and regress slightly on others compared with 1.0. When the language is specified, v1.5 is stronger than 1.0 on almost all supported languages. Set the tag when building the user message, for example processor.build_user_message(text=text_fr, language="French")."[pause 3.2s]". For example, ๆไปๅคฉๅญฆไน ไบไธ้ฆไธญๅฝ็ๅค่ฏ๏ผๅฎ็ๅๅญๆฏ[pause 3.2s]้ๅคๆ๏ผ inserts an explicit 3.2s pause before ้ๅคๆ.MOSS-TTS Local Transformer v1.5 supports 31 languages. It keeps the 20 languages supported by MOSS-TTS-Local-Transformer-v1.0 and extends multilingual continued training to additional languages including Cantonese, Dutch, Finnish, Hindi, Macedonian, Malay, Romanian, Swahili, Tagalog, Thai, and Vietnamese.
We recommend a clean, isolated Python environment with Transformers 5.0.0, or a recent Transformers version with Qwen3 support, to avoid dependency conflicts.
Install all required dependencies:
For better speed and lower GPU memory usage, you can install FlashAttention 2 if your hardware supports it.
If your machine has limited RAM and many CPU cores, you can cap build parallelism:
Notes:
pyproject.toml, which currently pins torch==2.9.1+cu128 and torchaudio==2.9.1+cu128.torch.float16 or torch.bfloat16.Tip: MOSS-TTS-Local-Transformer-v1.5 uses a fixed 12-codebook RVQ depth. Do not set
n_vq_for_inferenceto a value different fromconfig.n_vq.
MOSS-TTS-Local-Transformer-v1.5 provides the standard Hugging Face AutoProcessor and AutoModel interface. The examples below cover:
[pause X.Ys]trust_remote_code=True.processor.decode(...) returns audio tensors shaped as [channels, samples], so save them directly with torchaudio.save(path, audio, sampling_rate).OpenMOSS-Team/MOSS-Audio-Tokenizer-v2.sampling_rate to 48000 and n_vq to 12.You can serve MOSS-TTS-Local-Transformer-v1.5 with SGLang-Omni, which exposes an OpenAI-compatible /v1/audio/speech API for reference-less synthesis, zero-shot voice cloning, streaming, duration control, and language/style hints.
See the MOSS-TTS-Local cookbook for installation, full API details, deployment config, benchmarking, and limitations.
Install sglang-omni by following the SGLang-Omni installation guide, then download and serve the model:
A matching config file is available in SGLang-Omni at examples/configs/moss_tts_local.yaml.
Provide a reference clip and its transcript for better speaker similarity. audio_path may be a local path readable by the server, an HTTP(S) URL, or a base64 data URI.
ref_audio and ref_text are accepted as shorthand for references[0].audio_path and references[0].text.
Set "stream": true, "response_format": "pcm", and "stream_format": "audio" to receive raw 48 kHz PCM chunks. Pipe the stream through ffmpeg to write a playable WAV file:
Duration can be guided with an inline ${token:N} prefix or with token_count / duration_tokens. Inline markup such as [pause 0.5s], Pinyin, and IPA is passed through unchanged. Use language to hint the target language and instructions for free-form style guidance.
MOSS-TTS-Local-Transformer-v1.5 is API-compatible with MOSS-TTS-Local-Transformer-v1.0. For continuation with prefix audio, detailed UserMessage and AssistantMessage fields, generation hyperparameters, Pinyin/IPA preprocessing examples, and evaluation results, see the MOSS-TTS-Local-Transformer-v1.0.
If you use this model, please cite the MOSS-TTS Technical Report.
conda create -n moss-tts python=3.12 -yconda activate moss-ttsgit clone https://github.com/OpenMOSS/MOSS-TTS.gitcd MOSS-TTSpip install --extra-index-url https://download.pytorch.org/whl/cu128 -e ".[torch-runtime]"pip install --extra-index-url https://download.pytorch.org/whl/cu128 -e ".[flash-attn]" --no-build-isolationMAX_JOBS=4 pip install --extra-index-url https://download.pytorch.org/whl/cu128 -e ".[flash-attn]" --no-build-isolationfrom pathlib import Pathfrom tqdm import tqdmimport importlib.util
import torchimport torchaudiofrom transformers import AutoModel, AutoProcessor
# Disable the broken cuDNN SDPA backend on some CUDA/PyTorch combinations.torch.backends.cuda.enable_cudnn_sdp(False)# Keep these enabled as fallbacks.torch.backends.cuda.enable_flash_sdp(True)torch.backends.cuda.enable_mem_efficient_sdp(True)torch.backends.cuda.enable_math_sdp(True)
pretrained_model_name_or_path = "OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5"device = "cuda" if torch.cuda.is_available() else "cpu"dtype = torch.bfloat16 if device == "cuda" else torch.float32
def resolve_attn_implementation() -> str: # Prefer FlashAttention 2 when package + device conditions are met. if ( device == "cuda" and importlib.util.find_spec("flash_attn") is not None and dtype in {torch.float16, torch.bfloat16} ): major, _ = torch.cuda.get_device_capability() if major >= 8: return "flash_attention_2" # CUDA fallback: use PyTorch SDPA kernels. if device == "cuda": return "sdpa" # CPU fallback. return "eager"
attn_implementation = resolve_attn_implementation()print(f"[INFO] Using attn_implementation={attn_implementation}")
processor = AutoProcessor.from_pretrained( pretrained_model_name_or_path, trust_remote_code=True,)processor.audio_tokenizer = processor.audio_tokenizer.to(device)
text_zh = "ไบฒ็ฑ็ไฝ ๏ผๆฟไฝ ็ๆฏไธๅคฉ้ฝๅผๅพ่ขซ่ฎฐไฝ๏ผไนๅผๅพ่ขซ็ๆใ"text_en = "We stand on the threshold of the AI era, where intelligence becomes an extension of human creativity."text_fr = "Bonjour, je voudrais essayer une voix francaise naturelle et stable."text_pause = "ๆไปๅคฉๅญฆไน ไบไธ้ฆไธญๅฝ็ๅค่ฏ๏ผๅฎ็ๅๅญๆฏ[pause 3.2s]้ๅคๆ๏ผ"
# Use remote demo audio to avoid requiring local assets.ref_audio_zh = "https://speech-demo.oss-cn-shanghai.aliyuncs.com/moss_tts_demo/tts_readme_demo/reference_zh.wav"ref_audio_en = "https://speech-demo.oss-cn-shanghai.aliyuncs.com/moss_tts_demo/tts_readme_demo/reference_en.m4a"
conversations = [ # Direct TTS. Language tags are recommended in v1.5 when the language is known. [processor.build_user_message(text=text_zh, language="Chinese")], [processor.build_user_message(text=text_en, language="English")], [processor.build_user_message(text=text_fr, language="French")], # Explicit pause control. Use [pause X.Ys], such as [pause 3.2s]. [processor.build_user_message(text=text_pause, language="Chinese")], # Voice cloning with a reference audio. [processor.build_user_message(text=text_zh, reference=[ref_audio_zh], language="Chinese")], [processor.build_user_message(text=text_en, reference=[ref_audio_en], language="English")], # Duration control. At 12.5 frames per second, 125 frames is about 10 seconds. [processor.build_user_message(text=text_en, tokens=125, language="English")],]
model = AutoModel.from_pretrained( pretrained_model_name_or_path, trust_remote_code=True, attn_implementation=attn_implementation, torch_dtype=dtype,).to(device)model.eval()
batch_size = 1save_dir = Path("inference_root_moss_tts_local_v1_5")save_dir.mkdir(exist_ok=True, parents=True)sample_idx = 0
with torch.no_grad(): for start in tqdm(range(0, len(conversations), batch_size)): batch_conversations = conversations[start : start + batch_size] batch = processor(batch_conversations, mode="generation") input_ids = batch["input_ids"].to(device) attention_mask = batch["attention_mask"].to(device)
outputs = model.generate( input_ids=input_ids, attention_mask=attention_mask, max_new_tokens=4096, do_sample=True, audio_temperature=1.7, audio_top_p=0.8, audio_top_k=25, audio_repetition_penalty=1.0, )
for message in processor.decode(outputs): if message is None: continue audio = message.audio_codes_list[0] out_path = save_dir / f"sample{sample_idx}.wav" sample_idx += 1 # MOSS-TTS Local v1.5 codec returns stereo audio as [channels, samples]. # Save the two-channel tensor directly. torchaudio.save(str(out_path), audio, processor.model_config.sampling_rate)hf download OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5
sgl-omni serve \ --model-path OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5 \ --port 8000curl -X POST http://localhost:8000/v1/audio/speech \ -H "Content-Type: application/json" \ -d '{"input": "SGLang-Omni is a great project!"}' \ --output output.wavcurl -X POST http://localhost:8000/v1/audio/speech \ -H "Content-Type: application/json" \ -d '{ "input": "SGLang-Omni is a great project!", "references": [{ "audio_path": "https://huggingface.co/datasets/zhaochenyang20/seed-tts-eval-mini/resolve/main/en/prompt-wavs/common_voice_en_10119832.wav" }] }' \ --output output.wavcurl -N -X POST http://localhost:8000/v1/audio/speech \ -H "Content-Type: application/json" \ -d '{ "input": "Get the trust fund to the bank early.", "ref_audio": "https://huggingface.co/datasets/zhaochenyang20/seed-tts-eval-mini/resolve/main/en/prompt-wavs/common_voice_en_10119832.wav", "stream": true, "response_format": "pcm", "stream_format": "audio" }' \ | ffmpeg -f s16le -ar 48000 -ac 1 -i pipe:0 output_stream.wavcurl -X POST http://localhost:8000/v1/audio/speech \ -H "Content-Type: application/json" \ -d '{ "input": "${token:150}ไปๅคฉๅคฉๆฐไธ้ [pause 0.5s] ๅฐฑ่ฏฅๅบๅปๆๆๅคช้ณใ", "ref_audio": "https://huggingface.co/datasets/zhaochenyang20/seed-tts-eval-mini/resolve/main/en/prompt-wavs/common_voice_en_10119832.wav", "language": "Chinese" }' \ --output output_markup.wav