batch_whisper_extension

Batch Whisper (3 Speakers, Timestamped) — ComfyUI Node

A lightweight ComfyUI extension node that runs OpenAI Whisper to transcribe up to 3 audio inputs, merges and sorts timestamped segments, and outputs a single timestamped transcript string.

• Node class: __init__.BatchWhisperNode
• Implementation file: init.py
• Core helpers: __init__.WhisperModelWrapper, __init__.WhisperPatcher, __init__.WHISPER_PATCHER_CACHE, __init__.WHISPER_MODEL_SUBDIR

Installation

1. Drop the folder containing __init__.py into your ComfyUI extensions directory (e.g., ComfyUI/extensions/batch_whisper_extension).
2. Restart ComfyUI. The node will appear under the "Audio/Transcription" category as "Batch Whisper (3 Speakers, Timestamped)".

What this node does

• Accepts three audio inputs plus speaker name strings.
• Loads a Whisper model (selectable from the model list in the node) using a VRAM-safe patcher/cache.
• Saves each incoming waveform to a temporary WAV, transcribes with Whisper (word-level timestamps), and collects segments.
• Merges all segments, sorts by start time, and returns a timestamped transcript as a single STRING output.

Inputs

• audio1, audio2, audio3 (AUDIO): Audio objects produced by ComfyUI audio sources.
• speaker1, speaker2, speaker3 (STRING): Speaker labels to prefix each segment (defaults: Speaker1/2/3).
• model_name (enum): Whisper model selection; defaults to large-v3-turbo.
• language (optional): "auto" or a language selection; maps to Whisper language codes.
• prompt (optional STRING): Initial prompt passed to Whisper.

Output

• STRING: One combined transcript with speaker labels and timestamps formatted as: SpeakerName [start-end]: transcript text

Notes & behavior

• Model loading and offloading are handled by the VRAM-safe classes:
  • __init__.WhisperModelWrapper wraps a Whisper model instance and tracks weight memory.
  • __init__.WhisperPatcher manages loading/unloading and integrates with ComfyUI model management.
  • Patcher instances are cached in __init__.WHISPER_PATCHER_CACHE by model name to avoid repeated reloads.
• Temporary WAV files are written to the temp directory obtained from the included folder_paths helper. Temporary files are created per transcription call.
• If the Whisper model fails to load or transcription raises an exception for a particular input, the node injects an "[Error: ...]" segment for that speaker and continues.

Troubleshooting

• If you see memory errors, try a smaller model (e.g., small, base).
• If the node returns "Error: Whisper model failed to load.", ensure ComfyUI's device settings and PyTorch/CUDA are configured correctly.
• Check ComfyUI logs for messages produced by the node's logger.

References

• Node implementation: __init__.BatchWhisperNode — init.py
• Model wrapper/patcher: __init__.WhisperModelWrapper, __init__.WhisperPatcher
• Cache constant: __init__.WHISPER_PATCHER_CACHE
• Model subdirectory constant: __init__.WHISPER_MODEL_SUBDIR

License & credits

• Uses the open-source Whisper transcription code (whisper package). See upstream licenses for model and code terms.