<div align="center"> <a href="https://github.com/YanTianlong-01/comfyui_voicebridge"> <img src="./assets/Icon.png" width="200" height="200"> </a> <h1>VoiceBridge</h1> <p> <b>VoiceBridge translates spoken audio from any language to a target language while preserving the original speaker's voice characteristics, and generates bilingual SRT subtitle files.</b> </p> <br> </div>

🔄 Workflow

This node integrates ASR (Automatic Speech Recognition), LLM (Large Language Model), and TTS (Text-to-Speech) technologies to provide a complete speech translation pipeline.

Online Workflow

ASR + LLM + TTS Workflow

Universal TTS Workflow

✨ Features

🌍 Speech Translation: Convert speech from one language to any other language while retaining the original speaker's voice timbre.
🗣️ Multi-Language Support: Speech recognition and translation in dozens of languages covering all major global languages.
⏱️ Automatic Voice Alignment: The generated translated voice is automatically aligned with the original voice to stay in sync with the video content.
📝 Accurate Subtitle Generation: Through force-align technology, VoiceBridge produces accurate subtitles synchronized with the voice at the millisecond level.
🔌 Universal TTS Support: The new decoupled pipeline — Load SRT → SRT Splitter → <Any TTS node> → Audio List Merger by SRT — lets you drive any ComfyUI TTS custom node (Qwen3-TTS, VoxCPM, Fish Audio S2, LongCat-AudioDiT, CosyVoice, …). As long as the TTS node takes a STRING and returns an AUDIO.

📦 Installation

Via ComfyUI Manager (Recommended)

Search for "VoiceBridge" in ComfyUI Manager

Manual Installation

cd ComfyUI/custom_nodes
git clone https://github.com/YanTianlong-01/comfyui_voicebridge.git
cd comfyui_voicebridge
pip install -r requirements.txt

Model Directory Structure

VoiceBridge automatically searches or download models in the following priority:

ComfyUI/
├── models/
│   └── Qwen3-TTS/
│   |   ├── Qwen3-TTS-12Hz-1.7B-Base/
│   |   ├── Qwen3-TTS-12Hz-0.6B-Base/
│   └── Qwen3-ASR/
│       ├── Qwen3-ASR-1.7B/
│       ├── Qwen3-ASR-0.6B/
│       ├── Qwen3-ForcedAligner-0.6B/

or you can use the local model path. Setting local_model_path=qwen-tts/Qwen3-TTS-12Hz-1.7B-Base means use the model in ComfyUI/models/qwen-tts/Qwen3-TTS-12Hz-1.7B-Base.

For ASR models, you can set local_model_path_asr=qwen-asr/Qwen3-ASR-1.7B. For ForcedAligner models, you can set local_model_path_fa=qwen-asr/Qwen3-ForcedAligner-0.6B.

🧩 Nodes

📥 Load SRT

Loads an .srt (or .txt) subtitle file from ComfyUI/input/. The node ships with a dropdown listing existing files, plus a frontend-injected "choose srt file to upload" button that mirrors the UX of ComfyUI's built-in Load Audio node. Uploaded files are saved to the input directory and appear in the dropdown immediately.

Output is the raw SRT string, ready to be wired into SRT Splitter.srt_string.

| Input | Type | Description | |-------|------|-------------| | srt | dropdown | Pick an .srt / .txt file from ComfyUI/input/. Uploaded files are auto-added to this list. | | Output | Type | Description | | srt_string | STRING | Raw SRT content decoded from the file (UTF-8 with BOM / GBK fallback handled internally). |

✂️ SRT Splitter

Parses an SRT string and splits it into a list of per-sentence texts plus a metadata object that carries the original timestamps. Because texts is emitted as a native ComfyUI LIST, any downstream TTS node connected to it will be executed once per sentence automatically — no custom loop nodes required.

| Input | Type | Description | |-------|------|-------------| | srt_string | STRING | SRT text from Load SRT, Generate SRT, or a manual paste. | | Output | Type | Description | | texts | LIST[STRING] | List of per-sentence texts, one entry per SRT subtitle. ComfyUI's list-iteration will feed them to the downstream TTS one by one. | | srt_items | VB_SRT_ITEMS | Metadata object holding (index, start_ms, end_ms, text) for every subtitle entry. Feed this into Audio List Merger by SRT. | | count | INT | Number of subtitle entries parsed. |

🔀 Audio List Merger by SRT

Collects the per-sentence AUDIO outputs produced by any external TTS node (iterated over SRT Splitter.texts) and merges them back into a single AUDIO aligned with the original SRT timestamps. Uses the same duration-matching / tempo-limit / cascading-shift algorithm as the legacy SRT To Audio node. The sample rate is read from the first incoming AUDIO dict and every other sentence is resampled to match.

| Input | Type | Description | |-------|------|-------------| | srt_items | VB_SRT_ITEMS | Metadata from SRT Splitter carrying the original timestamps. | | audios | LIST[AUDIO] | Per-sentence AUDIO outputs from any TTS node — auto-collected by ComfyUI's list-iteration. | | tempo_limit | FLOAT | (optional) Maximum speed-up factor for audio that overflows a subtitle slot. Default 1.5, range 1.0–5.0. | | mini_gap_ms | INT | (optional) Minimum gap between consecutive subtitles in milliseconds. Default 100, range 0–10,000. | | Output | Type | Description | | audio | AUDIO | Final merged audio in ComfyUI format (waveform, sample_rate). | | adjusted_srt | STRING | Adjusted SRT string whose timing reflects the real audio durations after tempo / shift corrections. |

VoiceBridge ASR Loader

Loads the Qwen3-ASR model with auto-download support. This node manages model loading with intelligent caching and supports both HuggingFace and ModelScope model sources.

| Input | Type | Description | |-------|------|-------------| | repo_id | dropdown | Model repository selection (Qwen/Qwen3-ASR-1.7B or Qwen/Qwen3-ASR-0.6B) | | source | dropdown | Model download source: HuggingFace or ModelScope | | precision | dropdown | Computation precision: fp16, bf16, or fp32 | | attention | dropdown | Attention implementation: auto, flash_attention_2, sdpa, or eager | | max_new_tokens | INT | Maximum tokens to generate during transcription (default: 256, range: 1-4096) | | forced_aligner | dropdown | Optional forced aligner model for word-level timestamps (None or Qwen/Qwen3-ForcedAligner-0.6B) | | local_model_path_asr | STRING | Optional custom local path for ASR model loading | | local_model_path_fa | STRING | Optional custom local path for forced aligner model loading | | Output | Type | Description | | model | QWEN3_ASR_MODEL | Loaded ASR model instance for transcription operations |

VoiceBridge ASR Transcribe

Transcribes a single audio input to text using the loaded Qwen3-ASR model. Supports automatic language detection and optional timestamp generation for precise word-level timing.

| Input | Type | Description | |-------|------|-------------| | model | QWEN3_ASR_MODEL | Pre-loaded ASR model from VoiceBridge ASR Loader | | audio | AUDIO | Audio input in ComfyUI format (waveform and sample_rate) | | language | dropdown | Target language for transcription or "auto" for automatic detection | | context | STRING | Optional context hints to improve transcription accuracy for specialized content | | return_timestamps | BOOLEAN | Enable word-level timestamp output for precise timing information | | Output | Type | Description | | forced_aligns | LIST | List of ForcedAlignItem objects containing word-level timing data (if timestamps enabled) | | text | STRING | Transcribed text content from the audio input | | language | STRING | Detected or specified language code | | timestamps | STRING | Formatted string with word-level timestamps in "start_time-end_time: word" format (if enabled) |

Generate SRT

Generates an SRT subtitle file from transcribed text and forced alignment timestamps. This node combines text segmentation with timestamp data to create properly timed subtitle entries.

| Input | Type | Description | |-------|------|-------------| | forced_aligns | LIST | List of ForcedAlignItem objects with word-level timestamps from ASR transcription | | text | STRING | Input text content to convert to SRT format | | language | STRING | Language of SRT text | | save_srt | BOOLEAN | Enable automatic saving of the generated SRT file to output directory (default: True) | | file_name | STRING | Base file name for the SRT output (default: "VoiceBridge\subtitle") | | Output | Type | Description | | srt_string | STRING | Generated SRT formatted string with proper timestamps and segmentation |

VoiceBridge AI API

Provides integration with OpenAI-compatible API endpoints for AI-powered text processing. Supports customizable system prompts and generation parameters for flexible content generation workflows.

| Input | Type | Description | |-------|------|-------------| | model | STRING | API model identifier to use for text generation | | base_url | STRING | Base URL endpoint for the API (must be OpenAI-compatible) | | api_key | STRING | API authentication key for the service | | system_prompt | STRING | System instruction defining the AI assistant's behavior and constraints | | prompt | STRING | User prompt or query to send to the API | | max_tokens | INT | Maximum number of tokens in the generated response (default: 4096, range: 1-1000000) | | temperature | FLOAT | Sampling temperature controlling randomness (default: 0.7, range: 0-1) | | top_p | FLOAT | Nucleus sampling threshold for token selection (default: 0.95, range: 0-1) | | Output | Type | Description | | response | STRING | Generated text response from the API |

VoiceBridge TTS Loader

Loads the Qwen3-TTS text-to-speech model with auto-download support. This node handles model caching and supports multiple model variants for different voice synthesis needs.

| Input | Type | Description | |-------|------|-------------| | repo_id | dropdown | Model repository selection (Qwen/Qwen3-TTS-12Hz-1.7B-Base or Qwen/Qwen3-TTS-12Hz-0.6B-Base) | | source | dropdown | Model download source: HuggingFace or ModelScope | | precision | dropdown | Computation precision: fp16, bf16, or fp32 | | attention | dropdown | Attention implementation: auto, flash_attention_2, sdpa, or eager | | local_model_path | STRING | Optional custom local path for TTS model loading | | Output | Type | Description | | model | QWEN3_TTS_MODEL | Loaded TTS model instance for voice synthesis operations |

Voice Clone Prompt

Creates a voice clone prompt from reference audio for use with the Qwen3-TTS model. This enables voice cloning by extracting voice characteristics from a sample audio file.

| Input | Type | Description | |-------|------|-------------| | model | QWEN3_TTS_MODEL | Pre-loaded TTS model from VoiceBridge TTS Loader | | ref_audio | AUDIO | Reference audio sample for voice cloning (ComfyUI Audio format) | | ref_text | STRING | Transcript of the reference audio content (highly recommended for better quality) | | Output | Type | Description | | voice_clone_prompt | VOICE_CLONE_PROMPT | Voice clone prompt object containing extracted voice characteristics for TTS generation |

SRT To Audio

Converts SRT subtitle content to audio using the Qwen3-TTS model with voice cloning. Includes intelligent duration matching to align generated audio with subtitle timing, with optional speed adjustment for timing mismatches.

| Input | Type | Description | |-------|------|-------------| | model | QWEN3_TTS_MODEL | Pre-loaded TTS model from VoiceBridge TTS Loader | | srt_string | STRING | SRT formatted subtitle text to convert to audio | | voice_clone_prompt | VOICE_CLONE_PROMPT | Voice clone prompt from Voice Clone Prompt node | | language | dropdown | Target language for synthesis or "auto" for automatic detection | | tempo_limit | FLOAT | Maximum speed-up factor for audio that exceeds subtitle duration (default: 1.5, range: 1.0-5.0) | | mini_gap_ms | INT | Minimum gap between consecutive subtitles in milliseconds (default: 100, range: 0-10,000) | | batch_size | INT | Number of subtitle entries to process in each batch (default: 10, range: 1-1,000) | | Output | Type | Description | | audio | AUDIO | Generated audio in ComfyUI format (waveform and sample_rate) | | adjusted_srt | STRING | Adjusted SRT string with corrected timing based on actual audio durations |

Save SRT From String

Saves an SRT formatted string to a file in the ComfyUI output directory. Automatically handles file naming conflicts by appending incrementing indices.

| Input | Type | Description | |-------|------|-------------| | srt_string | STRING | SRT formatted string content to save | | file_name | STRING | Base file name for the SRT output (default: "VoiceBridge\subtitle") | | Output | Type | Description | | save_path | STRING | Full path where the SRT file was saved |

🧹 VoiceBridge Unload Model

Unloads cached ASR and/or TTS models from GPU memory to free VRAM. Useful when chaining different models sequentially or running on limited-memory GPUs. Acts as a pass-through node — wire the anything input after the node whose output you want to keep, and the same value will be forwarded on any once the unload runs.

| Input | Type | Description | |-------|------|-------------| | anything | * | Any value, passed straight through the node so you can insert it anywhere in a chain. | | Unload_ASR_Model | BOOLEAN | Unload the ASR model cache (default True). | | Unload_TTS_Model | BOOLEAN | Unload the TTS model cache (default True). | | Output | Type | Description | | any | * | Same value as anything (pass-through). |

🙏 Acknowledgments

Qwen3-ASR and Qwen3-TTS: Official open-source repository by Alibaba Qwen team.
ComfyUI-Qwen3-ASR: A nice and clean ComfyUI node by DarioFT.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.