ComfyUI Node
ACE-Step Transcriber
A ComfyUI node in ACE_STEP with 14 inputs and 1 output.
ACE-Step Transcriber
- audio
- transcription
◄model_idkana112233/ComfyUI-kaola-ace_step►
◄deviceauto►
◄dtypeauto►
◄languageauto►
◄chunk_length_s30►
◄return_timestampsfalse►
◄custom_prompt►
◄max_new_tokens4096►
◄temperature0.2►
◄top_p0.95►
◄repetition_penalty1.1►
◄num_beams1►
◄seed0►
CategoryACE_STEP
Inputs (14)
| Name | Type | Default | Description |
|---|---|---|---|
| audio | AUDIO | Input audio to transcribe. | |
| model_id | COMBO | kana112233/ComfyUI-kaola-ace_step | Select the Qwen2.5-Omni model. Can be a local path (in models/acestep-transcriber) or a HuggingFace ID. |
| device | COMBO | auto | Inference device. Use 'auto' or 'mps' for Mac. |
| dtype | COMBO | auto | Model precision. 'auto' uses float16 for CUDA and float32 for CPU/MPS. |
| language | COMBO | auto | Target language for transcription. 'auto' uses default prompt. |
| chunk_length_s | FLOAT | 300–300 | Audio chunk length in seconds for processing. |
| return_timestamps | COMBO | false | Whether to return timestamps. 'word' for word-level timestamps, 'true' for segment-level. |
| custom_prompt | STRING | Custom prompt to override built-in language prompts. e.g. 'Transcribe the audio to Chinese:' | |
| max_new_tokens | INT | 409664–8192 | Maximum number of tokens to generate. Increase for longer lyrics. |
| temperature | FLOAT | 0.20–1 | Sampling temperature. Lower values are more deterministic. |
| top_p | FLOAT | 0.950–1 | Nucleus sampling: cumulative probability threshold. |
| repetition_penalty | FLOAT | 1.11–2 | Penalty for repeating tokens. Increase if output gets stuck in loops. |
| num_beams | INT | 11–8 | Number of beams for beam search. 1 = no beam search. |
| seed | INT | 00–4294967295 | Random seed for reproducible results. 0 for random. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| transcription | STRING | — |