Load kotoba-whisper (Long-Form)

The long-form counterpart to kotoba-whisper's short-form path: this loads the model configuration built for transcribing audio that runs past Whisper's native ~30-second window, using a chunked ASR pipeline approach rather than a single forward pass.

How it works

Whisper's encoder is built around a fixed-length window; anything longer than that needs to be split into overlapping chunks, transcribed piece by piece, and stitched back together - that's what "long-form" means for any Whisper-family model, kotoba-whisper included. This loader configures that chunking pipeline (matching the parameters Hugging Face's own automatic-speech-recognition pipeline exposes for exactly this purpose) rather than the model weights alone.

The inputs and outputs that matter

• device - auto, cpu, or cuda. Same tradeoff as everywhere else in the pack: force cuda if you want a loud failure rather than a silent CPU fallback when benchmarking.
• chunk_length_s (default 15) - how long each processing chunk is, in seconds. Shorter chunks use less memory per step but mean more chunk boundaries for the pipeline to stitch across (a small risk of words getting split or repeated at seams); longer chunks are more memory-hungry but fewer stitch points.
• batch_size (default 16) - how many chunks get processed together in parallel. Higher uses more memory but transcribes faster on a GPU with room to spare; drop it if you're hitting memory limits on long files.
• model (output, KOTOBA_WHISPER_LONG) - feeds into whatever long-form transcribe node in this pack consumes it (not covered in this batch of articles, but it's the counterpart to SDT_KotobaWhisperTranscribeShort for clips over the short-form window).

How to install it

ComfyUI Manager: search ComfyUI-speech-dataset-toolkit, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kale4eat/ComfyUI-speech-dataset-toolkit
cd ComfyUI-speech-dataset-toolkit
pip install torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

Same as the short-form path: model weights download from Hugging Face on first use, so expect a real fetch the first time you load this.

Common issues & troubleshooting

Out of memory on a long file. Lower batch_size first - it's the more direct memory lever than chunk_length_s, since it controls how much runs concurrently rather than how the audio is divided.

Words repeated or dropped at what look like regular intervals. That's a classic sign of chunk-boundary stitching issues in long-form ASR generally - if it's bad enough to matter, try a larger chunk_length_s so there are fewer seams for the pipeline to reconcile.

This is genuinely obscure territory - kotoba-whisper's long-form path barely shows up anywhere outside its own Hugging Face model card, and this specific node wrapping it has essentially zero community footprint. If something behaves unexpectedly, the model card and the pack's source are your only real references; don't expect a Stack Overflow answer to exist.