MOSS-TTS-v1.5 8B — Voice Acting (24 kHz delay model)

A full fine-tune of OpenMOSS's MOSS-TTS-v1.5 (8B parameters, Qwen3-8B backbone, moss_tts_delay architecture, 32-codebook audio @ 24 kHz), specialised for expressive voice acting.

The fine-tune pushes the base model toward:

• Emotional delivery — happy, sad, angry, fearful, tender, sarcastic, whispered, shouted, and everything in between.
• Character voices — distinct personas driven from a natural-language instruction.
• Voice cloning — clone a speaker's timbre and delivery style from a short reference clip.

Sibling model: the 4.55-billion-parameter local-transformer variant at 48 kHz — laion/moss-tts-local-transformer-4.55b-voice-acting (merged, off-the-shelf).

It keeps the multilingual coverage of the base model (English, German, Chinese, and many more — see the language list), and is fully API-compatible with MOSS-TTS-v1.5: same processor, same generate interface, same input schema.

---

Model at a glance

---

Prompting

The model is prompted through the processor's native build_user_message fields:

• instruction — how to say it: the emotion, character, or performance direction (e.g. "Speak like a weary old sailor telling a ghost story, low and gravelly.").
• text — what to say: the words to be spoken.
• language — the language of text (e.g. "English", "German", "Chinese"). Set it whenever the language is known; it improves multilingual stability.
• reference — optional audio (a file path, URL, or waveform tensor) to clone a voice. Omit it to let the model invent a voice consistent with the instruction.

---

Usage

from transformers import AutoModel, AutoProcessor
import torch, torchaudio

# Recommended SDPA backend settings (the cuDNN SDPA path is broken for this model)
torch.backends.cuda.enable_cudnn_sdp(False)
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_mem_efficient_sdp(True)
torch.backends.cuda.enable_math_sdp(True)

model_id = "laion/moss-tts-v1.5-8b-voice-acting"
device, dtype = "cuda", torch.bfloat16

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
processor.audio_tokenizer = processor.audio_tokenizer.to(device)

model = AutoModel.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=dtype,
    attn_implementation="sdpa",   # or "flash_attention_2" if installed
).to(device).eval()

def synth(message, out_path):
    batch = processor([[message]], mode="generation")
    with torch.no_grad():
        outputs = model.generate(
            input_ids=batch["input_ids"].to(device),
            attention_mask=batch["attention_mask"].to(device),
            max_new_tokens=4096,
        )
    audio = processor.decode(outputs)[0].audio_codes_list[0]
    torchaudio.save(out_path, audio.unsqueeze(0).cpu().float(),
                    processor.model_config.sampling_rate)

# 1) Expressive delivery, model invents a voice (no reference)
synth(processor.build_user_message(
    instruction="Excited and breathless, like sharing amazing news with a best friend!",
    text="You will not believe what just happened. We actually did it!",
    language="English",
), "invented_voice.wav")

# 2) Voice cloning — clone the timbre/style of a reference clip
synth(processor.build_user_message(
    instruction="Calm, warm, reassuring bedtime-story narrator.",
    text="Once upon a time, in a quiet little village, everyone slept soundly.",
    language="English",
    reference=["/path/to/reference.wav"],
), "cloned_voice.wav")

processor.decode(...) returns a list of messages; each carries the decoded waveform in message.audio_codes_list[0] at processor.model_config.sampling_rate (24 kHz).

Sampling settings. A grid search over temperature and audio repetition penalty (scored for intelligibility, quality, and genuineness on a held-out English + German set) gives good defaults of temperature 1.0, repetition penalty 1.1 with a reference clip, and temperature 0.8, repetition penalty 1.1 without a reference. See GENERATION_SETTINGS.md for the full results table and per-metric best configs.

---

Training

This model was fine-tuned on a mix of the Emolia corpus and synthetic speech generated by several TTS models — primarily Gemini-3.1 TTS and DramaBox TTS. The goal of the mix was to concentrate on expressive, emotional, and character-driven delivery while preserving the base model's multilingual and voice-cloning abilities. It is a conservative full fine-tune: the base model's general capabilities are largely retained.

---

Tips

• Put performance cues in instruction. Emotion, character, pacing, and vocal-burst / performance cues (laughter, sighs, gasps, whispering, shouting) belong in instruction, not in text.
• German text: write it naturally. Use lowercase with proper umlauts (ä ö ü, not ae/oe/ue) and avoid ALL-CAPS, which the model tends to mispronounce. Reserve capitals for normal sentence casing.
• Use punctuation for emphasis and prosody. !, ?, ., and ... meaningfully shape intonation and pauses — lean on them instead of capitalisation.
• Set language whenever you know it; it stabilises multilingual synthesis.
• Data augmentation. Voice cloning combined with text paraphrase is an effective way to generate additional, style-consistent training material.

---

Limitations

• This is a conservative full fine-tune. It broadens expressive and voice-acting behaviour, but it may not beat task-specific baselines on every axis (e.g. raw intelligibility, speaker similarity, or a single narrow language/domain).
• Expressiveness is driven by the instruction field; vague or contradictory instructions can produce inconsistent delivery. Iterate on wording.
• As with any generative TTS, outputs should not be used to impersonate real individuals without consent. Use responsibly.

---

License

Released under the Apache-2.0 license.