dessix/moss-ttsd 🔢📝✓🖼️ → 🖼️

▶️ 694 runs 📅 Jul 2025 ⚙️ Cog 0.16.0 🔗 GitHub ⚖️ License

dialogue-tts spoken-dialogue text-to-speech voice-cloning

About

MOSS-TTSD (text to spoken dialogue) is an open-source bilingual spoken dialogue synthesis model that supports both Chinese and English. It can transform dialogue scripts between two speakers into natural, expressive conversational speech.

Example Output

Output

Performance Metrics

22.04s Prediction Time

22.06s Total Time

All Input Parameters

{
  "seed": 42,
  "text": "[S1]诶，跟你说个事儿啊，我最近听了不少那种AI生成的播客，不知道你有没有听过。[S2]哦，听过一些。怎么了，感觉怎么样？[S1]就是……怎么说呢，单听一句话，你觉得，哇，好像跟真人没啥区别。[S2]嗯。[S1]但是，你只要让它说上一段完整的对话，比如俩人聊天那种，那个感觉就立马不对了。[S2]对对对，我懂你的意思。就是那个所谓的“恐怖谷”效应，是吧？听着有点瘆人，感觉特别假，没有那个交流感。[S1]就是这个词儿，恐怖谷。结果你猜怎么着，这个事儿最近好像有救了。",
  "use_normalize": true,
  "reference_text_speaker1": "周一到周五每天早晨七点半到九点半的直播片段,言下之意呢就是废话有点多,大家也别嫌弃,因为这都是直播间最真实的状态了",
  "reference_text_speaker2": "如果大家想听到更丰富更及时的直播内容,记得在周一到周五准时进入直播间,和大家一起畅聊新消费新科技新趋势",
  "reference_audio_speaker1": "https://replicate.delivery/pbxt/NQ5EtOr3vU5lQdCgXwDM5ywZRwEez5LYv7Mx3XOHHe8vWqQB/zh_spk1_moon.wav",
  "reference_audio_speaker2": "https://replicate.delivery/pbxt/NQ5Eu4HdpF1khhnx36ivEFa54pxlpBHg7Zx8RLrzIxlEutKx/zh_spk2_moon.wav",
  "reference_audio_speaker1_base64": "",
  "reference_audio_speaker2_base64": ""
}

Input Parameters

seed Type: integerDefault: 0: Random seed for reproducibility
text Type: stringDefault: [S1]你好[S2]你好，最近怎么样[S1]还不错，你呢[S2]我也挺好的，谢谢关心: Dialogue text, format: [S1]Speaker 1 content[S2]Speaker 2 content[S1]...
use_normalize Type: booleanDefault: true: Whether to use text normalization (recommended for better handling of numbers, punctuation, etc.)
reference_text_speaker1 Type: stringDefault: 周一到周五每天早晨七点半到九点半的直播片段,言下之意呢就是废话有点多,大家也别嫌弃,因为这都是直播间最真实的状态了: Reference text for speaker 1 (corresponding to reference audio)
reference_text_speaker2 Type: stringDefault: 如果大家想听到更丰富更及时的直播内容,记得在周一到周五准时进入直播间,和大家一起畅聊新消费新科技新趋势: Reference text for speaker 2 (corresponding to reference audio)
reference_audio_speaker1 Type: string: Reference audio file for speaker 1 (optional, for voice cloning)
reference_audio_speaker2 Type: string: 说话者2的参考音频文件（可选，用于声音克隆）/ Reference audio file for speaker 2 (optional, for voice cloning)

Output Schema

Output

Type: string • Format: uri

Example Execution Logs

Processing text: [S1]诶，跟你说个事儿啊，我最近听了不少那种AI生成的播客，不知道你有没有听过。[S2]哦，听过一些。怎么了，感觉怎么样？[S1]就是……怎么说呢，单听一句话，你觉得，哇，好像跟真人没啥区别。[S2...
Using voice cloning with reference audio
Processing 1 samples starting from index 0...
Using speaker1 and speaker2 information for prompt audio and text.
Starting batch audio generation...
Original outputs shape: torch.Size([1, 823, 8])
Start value: 467
Shape after slicing: torch.Size([1, 356, 8])
MAX_CHANNELS: 8
Calculated seq_len: 816
Speech token shape for sample 0: torch.Size([349, 8])
Audio generation completed: sample 0
Successfully generated audio with duration: 27.92 seconds

Version Details

Version ID: 45485ef4ee8ad08ee718ed2ae59d4c8f1cb682ad7842f1a31092c6fbe7c575f5
Version Created: August 13, 2025

Run on Replicate →