ComfyUI Node

TS Whisper

Transcribe speech to text, with optional SRT subtitles, on the local OpenAI Whisper engine.

By AlexYez·Created 2 years ago·Updated 3 days ago· 11
TS Whisper
  • audio
  • srt_content
  • text_content
  • ttml_content
modellarge-v3
tasktranscribe
source_languageru
timestampssegment
precisionfp16
beam_size5
temperature0.0
temperature_fallbacks0.0,0.2,0.4,0.6,0.8,1.0
condition_on_previous_texttrue
compression_ratio_threshold2.4
logprob_threshold-1.0
no_speech_threshold0.60
initial_prompt
save_srt_filetrue
output_filename_prefixtranscribed_audio
output_dir/tmp/ComfyUI/output
CategoryTS/Audio

Inputs (17)

NameTypeDefaultDescription
audioAUDIOComfyUI audio input (waveform + sample_rate).
modelCOMBOlarge-v3Whisper model: large-v3 (best quality) or turbo (faster). Weights and cache are shared with TS Super Prompt.
taskCOMBOtranscribetranscribe = recognize in the source language; translate_to_english = translate to English (turbo cannot translate and falls back to transcription).
source_languageCOMBOruLanguage of the input audio. auto = auto-detect. Use ru for Russian.
timestampsCOMBOsegmentsegment = per-segment timecodes; word = per-word; none = no timecodes (SRT/TTML disabled).
precisionCOMBOfp16fp16 is faster on GPU; fp32 is more stable. CPU always uses fp32.
beam_sizeINT51–10Beam search width (used when temperature=0). 1 = greedy. Higher = more accurate but slower.
temperatureFLOAT0.00–1Decoding temperature. 0 = deterministic.
temperature_fallbacksSTRING0.0,0.2,0.4,0.6,0.8,1.0Comma-separated temperature ladder (e.g. 0.0,0.2,0.4,0.6,0.8,1.0). Overrides temperature.
condition_on_previous_textBOOLEANtrueFeed context from previous segments (more coherent, but can amplify drift).
compression_ratio_thresholdFLOAT2.40–10zlib compression threshold for the hallucination detector (whisper default 2.4).
logprob_thresholdFLOAT-1.0-10–0Average log-probability threshold (whisper default -1.0).
no_speech_thresholdFLOAT0.600–1No-speech probability threshold (whisper default 0.6).
initial_promptSTRINGInitial prompt to bias vocabulary, names, and style.
save_srt_fileBOOLEANtrueSave SRT/TTML files to the output/subtitles folder.
output_filename_prefixSTRINGtranscribed_audioFilename prefix for SRT/TTML.
output_diroptSTRING/tmp/ComfyUI/outputOutput folder for SRT/TTML. Empty = default output directory.

Outputs (3)

NameTypeDescription
srt_contentSTRINGSubtitles in SRT format (empty when timestamps=none).
text_contentSTRINGFull transcription as plain text.
ttml_contentSTRINGSubtitles in TTML format (empty when timestamps=none).