TS Whisper

The name is a lie in the best way: it doesn't call any API and needs no key. TS Whisper is speech-to-text on the native OpenAI Whisper engine, running entirely on your machine, and it hands you subtitles in three formats at once. Wire in an AUDIO clip - from a loader, from the pack's audio loader, from anywhere - and out come SRT, plain text, and TTML. Transcribe a voiceover, generate subtitles for a video you just saved, or pull the text out of a podcast before feeding it to an LLM node. It's the kind of node you reach for constantly once it's in your graph.

How it works

Under the hood is a shared engine (nodes/_whisper_engine.py) that this node and TS Super Prompt's voice input both use - same weights, same on-disk folder (models/whisper/), and an in-memory model cache so loading large-v3 here and again for Super Prompt voice reuses the same model object instead of holding two copies in memory. That shared-engine detail is why bug fixes land in both nodes at once.

The two model choices are simple: large-v3 (default) for best quality, or turbo (the distilled large-v3-turbo) for speed. Weights download on first use into models/whisper/.

The inputs that matter

• audio - a ComfyUI AUDIO input (waveform + sample rate).
• model - large-v3 or turbo.
• task - transcribe (recognize in the source language) or translate_to_english. One gotcha: turbo cannot translate and falls back to transcription, so if you need English out of Russian audio, use large-v3.
• source_language - auto for detection, or set it (ru for Russian) - setting it speeds things up and avoids misdetection.
• timestamps - segment (per-segment timecodes), word, or none. With none, the SRT/TTML outputs come back empty; the plain text always works.
• precision - fp16 on GPU, fp32 for stability. CPU always uses fp32. And on a Mac this node always runs on CPU - openai-whisper doesn't run reliably on Metal, so the option is just not offered rather than offered and broken.
• initial_prompt - a genuinely useful field: it biases vocabulary, names and style. Feed it a list of proper nouns from your domain and watch the spelling stop mangling them.
• save_srt_file / output_filename_prefix - whether to also write SRT/TTML files to output/subtitles, and under what name.

The deeper decoding controls - beam_size (1 is greedy, higher is more accurate but slower), temperature and the temperature_fallbacks ladder, condition_on_previous_text, and the three hallucination thresholds (compression_ratio_threshold, logprob_threshold, no_speech_threshold) - are Whisper's standard dials, exposed with sensible defaults. Leave them alone until you hit a problem; the threshold trio is what you tune when the model starts inventing words on silence.

The outputs are srt_content, text_content, and ttml_content - all STRING, all wireable straight into save nodes, note nodes, or an LLM.

Installing it

Part of comfyui-timesaver: ComfyUI Manager → search Timesaver, or

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt

then restart ComfyUI. openai-whisper is in the base requirements, so a plain install covers it. No system ffmpeg needed - the pack ships imageio-ffmpeg's static binary, and the node asks for that binary before ever looking at your PATH.

Gotchas

If transcription fails with an ffmpeg-ish error, it's the dependency, not you: python -m pip install --upgrade imageio-ffmpeg (with the same Python ComfyUI runs from). If you're getting hallucinations on silence, that's the thresholds - raise no_speech_threshold a bit, and feed an initial_prompt for domain names. And if you're on a Mac, don't chase GPU speed here; the node deliberately runs on CPU because Metal is unreliable with this engine, and it says so instead of pretending otherwise.