BytePlus Seed Speech ASR

Transcription is the least glamorous node in this pack and the one that quietly enables the rest of it. The Podcast Clip template writes a two-voice script, synthesises it, then runs this node to produce SRT subtitles. The Multilingual Dubbing template transcribes a speech clip, translates it with the LLM, and re-voices it. Neither works without ASR, and if you've tried to get word-level timings out of a local Whisper node, you know why it's nice to have it as one call.

It's also the node where the fast/standard split matters most, because that split decides how your audio gets to the service.

How it works

model picks the path:

• seed-asr-fast (default) sends a connected clip inline - one request, up to 2 hours or 100 MB. No upload, no login, done.
• seed-asr-2.0 / seed-asr-1.0 submit and poll, handle up to 5 hours and more languages - but they only accept URLs. So a connected clip has to be uploaded to Comfy.org storage first, which means a Comfy.org login, or you supply a public URL.

Audio comes in via audio (a Load Audio node, for instance) or audio_url; one or the other. audio_format tells the service what container your URL is, and auto reads the extension - set it explicitly for URLs that don't have one.

The inputs worth setting

• language defaults to auto, which covers Chinese, English and Chinese dialects. For anything else, name the language - and note the last sixteen languages on the list require the 2.x/1.0 models, not fast.
• enable_punc (punctuation), enable_itn (numbers as digits) and enable_ddc (drop filler words and repetitions) are the readability trio; ddc is the one that turns "um, so, basically" into a usable transcript.
• enable_speaker_info labels speakers - best with ten or fewer people, per the tooltip. enable_lid tags each utterance with its detected language, enable_channel_split treats stereo left and right as separate speakers, and enable_auto_lang detects the spoken language, overriding language.
• hotwords is the accuracy lever for names, brands and jargon: one per line or comma-separated, up to 5000, and it's Chinese-English model only (auto or zh-CN).
• context_text lets you give the model dialogue history or a scene description so it stops guessing at homophones, and context_image_url / context_image (ASR 2.0) gives it a picture for visual context - the pack resizes the connected image to fit the 500 KB limit before uploading it.
• Segmentation: vad_segment splits on silence instead of meaning, end_window_size sets the silence in milliseconds that ends a sentence (300–5000, 0 = semantic), and output_zh_variant converts Chinese output to Traditional (traditional, tw, hk).
• Cleanup: remove_words deletes terms, mask_words replaces them with *, filter_system_sensitive_words masks the built-in list, and wrap_sensitive_words wraps filtered words in backticks instead of hiding them - that last one is what you want if you need to see what was censored rather than silently lose it.

Outputs are text (the plain transcript), utterances_json (utterances with timings), srt (subtitle file contents, ready for a Save Text) and duration as a float.

Install and the key

cd ComfyUI/custom_nodes
git clone https://github.com/byteplus-sa/ComfyUI-BytePlus-ModelArk
pip install -r ComfyUI-BytePlus-ModelArk/requirements.txt

Restart (ComfyUI 0.31.0 or newer), or install from Manager by searching BytePlus ModelArk. Seed Speech has its own API key - same row in Settings → BytePlus, or BYTEPLUS_SEED_SPEECH_API_KEY in user/.env - and it runs in ap-southeast-1 only. ModelArk keys don't work here; you'll get Invalid X-Api-Key if you try.

Where people get burned

The login requirement on the standard models is the number one surprise. You connect Load Audio, run, and get told a Comfy.org login is needed - because seed-asr-2.0 and 1.0 only take URLs and your clip had to be uploaded somewhere first. seed-asr-fast avoids the whole problem; the tradeoff is the language list and the two-hour ceiling.

The other one is the misunderstanding about what ASR is for here. This isn't a captioning tool for describing images, and it isn't a diarisation service for a noisy twelve-person meeting - speaker labelling works best with ten or fewer clean voices, and heavy crosstalk is where it falls down like everything else. Feed it a voiceover, a podcast, a dubbed clip, and it does the job in one call, SRT included.