ASR · 语音转字幕

If you've ever fought Whisper's VRAM footprint, or wanted speaker diarization without stitching three tools together, this is the easy button. ASR · 语音转字幕 is the cloud speech-recognition node: upload a video, get back timestamped subtitle text, no GPU required. Given it runs on Volcengine's AI MediaKit, it's also a ByteDance-adjacent stack that happens to be genuinely strong at Mandarin.

What you set

• content_type - auto, speech (dialogue) or singing (vocals). auto is the sensible default; pick singing for music-heavy clips so it doesn't try to force song lyrics into prose.
• language - auto, cmn-Hans-CN (Mandarin) or eng-US. Note the list is short: auto plus Chinese and US English. For any other language you're trusting auto-detection.
• enable_speaker_info (BOOLEAN, default false) - turns on speaker diarization, i.e. "who said this" labels. This is the feature that usually costs you a separate pipeline locally.
• enable_confidence (BOOLEAN, default false) - adds per-segment confidence scores.
• Plus the usual video, poll_interval_seconds (default 10) and max_poll_attempts (default 720).

Two outputs: subtitle - the timestamped subtitle text as a STRING - and raw_json, the full result with keys and signed URLs redacted. Wire subtitle into a subtitle-burn node or a script that writes an SRT, or parse raw_json if you need per-segment detail.

How it works

Same cloud choreography as every node here: the video becomes a temp file, mediakit-cli submits an async recognition task to your MediaKit account, the node polls, then downloads the subtitle text file and decodes it (UTF-8 with a GB18030 fallback, which tells you who the target audience is). The result comes back as plain text you can feed straight downstream.

Install and the cloud setup

ComfyUI Manager, search "MediaKit Toolkit" / mediakit-toolkit, full restart - or:

cd ComfyUI/custom_nodes
git clone https://github.com/fudanll-sys/comfyui-mediakit-toolkit.git

No Python deps, no models. You need a Volcengine account with AI MediaKit activated, an API key, and the CLI authed in the environment that launches ComfyUI:

npx @volcengine/mediakit-cli install -y
mediakit-cli init --mode cloud-first --api-key "YOUR_MEDIAKIT_API_KEY" --credential-store config --yes

Running the node submits a paid cloud task and uploads your audio to Volcengine - that's the trade you're making.

The honest comparison

Local faster-whisper is free and will transcribe a short clip on CPU. If you just need plain transcription of English audio, stay local. Where this node earns its keep is the combination: zero setup, built-in speaker diarization, confidence scores, and a cloud that handles Chinese dialogue notably well. The failure mode to watch is the one every API node shares - it's metered, and long videos add up. Transcribe a minute first, sanity-check the timestamps, then commit the whole episode.