✏️ ASR Transcribe

This is the suite running in reverse. Every other node here takes text and makes speech; ASR Transcribe takes speech and makes text. Feed it an audio clip and it gives you a transcript plus timing data - the raw material for subtitles, for cleaning up an existing recording, or for grabbing the exact words of a reference voice clip so you can hand them to a cloning engine.

It's a consumer node like TTS SRT: it needs an ASR-capable engine wired in. In this suite that means a Qwen3-TTS engine (which does double duty as Qwen3-ASR) or a Granite ASR engine. This node holds the transcription settings; the engine holds the model.

How it works

You connect the audio and the engine, pick a language (or Auto), and it runs the model over the clip. Long audio is handled by chunk_size with an overlap between chunks so words at the boundaries aren't lost. The node's real value in this pack is that it doesn't just dump text - it emits structured timing data, which the 📺 Text to SRT Builder can turn into properly-timed subtitles. That split (transcribe here, build subtitles downstream) is deliberate: you can clean or edit the transcript in between and still keep the original timings.

The inputs and outputs that matter

• engine - required. A Qwen3 or Granite ASR engine node. No engine, no transcription.
• audio - the clip to transcribe.
• language - Auto usually works, but pin it if auto-detect guesses wrong on a short or accented clip.
• task - transcribe (same language) or translate (into the engine's target language).
• timestamps - none or word. Choose word when you want word-level timing for tight subtitles.
• diarization - speaker labels (who said what). Note this depends on the engine: the Granite 4.1 plus model does native diarization; not every engine will.

chunk_size / overlap tune long-audio handling, and enable_asr_cache skips re-transcribing unchanged audio.

Three outputs: text (the plain transcript), asr_timing_data (the structured timings - wire this into Text to SRT Builder), and info (diagnostics). Leave asr_timing_data disconnected and the SRT builder will estimate timings from text alone.

Installing it

Ships with the pack. ComfyUI Manager → "TTS Audio Suite" → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/diodiogod/TTS-Audio-Suite.git
cd TTS-Audio-Suite
python install.py

The ASR model weights (Qwen3-ASR or Granite, several GB) auto-download on first use through whichever engine you connect.

Common issues & troubleshooting

Nothing transcribes. The engine input needs to be an ASR-capable engine - a Qwen3-TTS or Granite ASR node. A pure TTS engine won't do it.

Diarization is empty or ignored. Speaker attribution isn't universal. If you need who-said-what, use a Granite engine that supports it (the 4.1 plus variant), and be aware that asking for diarization and word timestamps together makes the suite fall back to the reused Qwen forced aligner to deliver both.

Words dropped at chunk boundaries on long audio. That's what overlap is for. Bump it up so consecutive chunks share more context and stop clipping words at the seams.

Wrong language detected. Auto can misfire on short clips. Set language explicitly. And if you actually wanted a translation rather than a same-language transcript, switch task to translate.