Seed Speech ASR

Node name: BytePlus Seed Speech ASR
Node ID: BytePlusSeedASR
Find it: Double-click the canvas and search for “Seed Speech ASR”, or browse BytePlus ModelArk/Speech
Keys needed: Seed Speech API key
Runs when: Its output is used by another node, such as Save Image or Save Video. It is not an output node.

What it does

Turns spoken audio into text. You connect an audio clip, for example from Load Audio, or paste a public audio link, and you get the transcript, the timing of each utterance and ready-made SRT subtitles.

The node needs your separate Seed Speech API key (Singapore region only), bills every run and runs only when another node uses one of its outputs. seed-asr-fast answers in one request, while seed-asr-2.0 and seed-asr-1.0 submit a task and wait for it.

On the canvas

The Seed Speech ASR node with optional audio and context_image sockets, model seed-asr-fast, an empty audio_url and the outputs text, utterances_json, srt and duration — Default settings. Connect audio, or fill audio_url.

Inputs

Input	Type	Default	What it does
`model`	choice	`seed-asr-fast`	seed-asr-fast: one request, audio up to 2 h / 100 MB. seed-asr-2.0 / 1.0: submit and poll, up to 5 h, more languages; connected audio is uploaded to Comfy.org storage.Choices: seed-asr-fast, seed-asr-2.0, seed-asr-1.0
`audio_url`	text		Public http(s) audio URL. Use this or the audio input (e.g. Load Audio).
`language`	choice	`auto`	auto recognizes Chinese, English and Chinese dialects; pick a language for others. The last 16 languages need seed-asr-2.0 or 1.0.Choices: auto, en-US, zh-CN, yue-CN, ja-JP, ko-KR, id-ID, ms-MY, th-TH, vi-VN, fil-PH, hi-IN … (56 in total)
`enable_punc`	on/off	`on`	Add punctuation.
`enable_itn`	on/off	`on`	Write numbers in digits (inverse text normalization).
`enable_ddc`	on/off	`off`	Remove filler words and repetitions.
`enable_speaker_info`	on/off	`off`	Label speakers (best with 10 or fewer).
`hotwords`	text		Names or terms to favour, one per line or comma-separated (up to 5000). Chinese-English model only: language auto or zh-CN.
`audio`optional	AUDIO		Audio to transcribe, e.g. from Load Audio.
`context_image`optional	IMAGE		ASR 2.0 visual context image (uploaded to Comfy.org storage, resized to fit 500 KB).

Advanced inputs (13)

Input	Type	Default	What it does
`context_text`	text		Dialogue history or scene description, one entry per line, newest first (up to 20 entries / 800 tokens). Prefix a line with 'user:' or 'bot:' to mark the speaker. Chinese-English model only.
`context_image_url`	text		ASR 2.0 visual context: public JPEG/PNG URL (up to 500 KB). Or connect context_image.
`enable_auto_lang`	on/off	`off`	Detect the spoken language automatically (overrides language).
`enable_lid`	on/off	`off`	seed-asr-2.0 / 1.0: add a detected-language label (lid_lang) to each utterance.
`enable_channel_split`	on/off	`off`	Recognize the left and right channels of stereo audio separately (channel_id 1 / 2).
`vad_segment`	on/off	`off`	Split sentences on silence (VAD) instead of meaning.
`end_window_size`	integer	`0`	Silence in ms that ends a sentence (300-5000; 0 = semantic segmentation).Range: 0–5000
`output_zh_variant`	choice	`none`	Convert Chinese output to Traditional Chinese: traditional, tw (Taiwan) or hk (Hong Kong).Choices: none, traditional, tw, hk
`filter_system_sensitive_words`	on/off	`off`	Mask words from the built-in sensitive word list with *.
`remove_words`	text		Words to delete from the transcript, comma-separated.
`mask_words`	text		Words to replace with * in the transcript, comma-separated.
`wrap_sensitive_words`	on/off	`off`	Wrap filtered words in backticks instead of hiding them silently.
`audio_format`	choice	`auto`	Container of audio_url. auto reads the file extension; the standard models read the file itself when the link has none. Set it if detection fails.Choices: auto, wav, mp3, ogg, m4a, aac, amr, spx, pcm, raw

Notes on inputs

Audio source: connect audio or fill audio_url, never both. audio_url must be a public http(s) link.
Fast or standard: with seed-asr-fast, a connected clip is sent inside the request. The standard models take audio only as a link, so the node first uploads a connected clip to Comfy.org storage. That needs a Comfy.org login in ComfyUI.
Language: auto recognizes Chinese, English and Chinese dialects. Choose a language for anything else. The last 16 languages in the list work only with seed-asr-2.0 and seed-asr-1.0.
Hotwords and context: hotwords and context_text work with the Chinese-English model only, so keep language on auto or zh-CN and enable_auto_lang off. context_image and context_image_url give a visual hint (an ASR 2.0 feature that seed-asr-1.0 refuses). A connected context_image is uploaded to Comfy.org storage, so it needs a Comfy.org login; a context_image_url is sent as it is.
audio_format: with auto the node reads the extension of the link. For a link without one, it looks at the first bytes of the file, in standard mode. Set the format yourself if that does not work.
Speakers and channels: enable_speaker_info and enable_channel_split add speaker or channel labels to srt. Channel split needs stereo audio.

Outputs

Output	Type	What it holds
`text`	text
`utterances_json`	text
`srt`	text
`duration`	number

Notes on outputs

utterances_json holds the utterances the service returns, with their timings. srt is built from them.
duration is the audio length in seconds, as reported by the service.
If the service finds no speech, the node does not fail: the outputs are empty (duration 0) and the console says “No speech was found in the audio.”

Limits and tips

seed-asr-fast takes audio up to 2 hours and 100 MB. A connected clip is converted to 16 kHz WAV before it is sent, and that WAV must stay within 100 MB.
seed-asr-2.0 and seed-asr-1.0 take audio up to 5 hours. The node checks the task every 2 seconds and waits for up to 12,600 seconds.
Uploads to Comfy.org storage are cached for 12 hours, so running the same clip again does not upload it again.
seed-asr-fast accepts raw, wav, mp3 and ogg audio from a link. Use seed-asr-2.0 for m4a, aac, amr, spx or pcm.
enable_lid (language labels) is for the standard models. end_window_size is 0 or between 300 and 5000 ms.
Interrupting a run stops the node’s wait. The node sends no cancel request, so a task that was already submitted may still be billed.

Common errors

Message	What to do
`[BytePlus] Connect audio or fill audio_url.`	Give the node one audio source.
`[BytePlus] Use audio or audio_url, not both.`	Disconnect the `audio` socket or empty `audio_url`.
`[BytePlus] Uploading the audio to Comfy.org storage failed: …`	Log in to your Comfy.org account in ComfyUI, or use a public `audio_url` instead.
`[BytePlus] Set audio_format: the audio_url has no recognizable extension, …`	Choose the format in `audio_format`, or use a link that ends in the file extension.
`[BytePlus] Hotwords and context only work with the Chinese-English model: …`	Set `language` to `auto` or `zh-CN`, and turn off `enable_auto_lang`.