FireRedAudio 逐句制作审核台 · T8star-Aix

Automated QA can catch a clipped peak, a wrong word, an overrun timecode. It cannot tell you whether the performance is good - that's a human ear, and it's the last thing between your batch and delivery. T8_FireRedAudio_LineReview is that gate, built as a node: an in-node table where you audition each line, see what QA flagged, and give it a verdict. Only the lines you approve make it into the delivery batch. Everything else stays out, explicitly marked, so it can't silently sneak into an export.

What it does

You feed it the production audio_batch and, optionally, a SpeechQA object to pre-populate the table with QA suggestions. The node's front end shows a review board - per line, you can listen, download, see QA flags, then assign a decision: approve, mark for human review, or redo. You also give a 1–5 rating and can add notes. That human judgment is written back into the node's serializable inputs (decisions_json, ratings_json, notes_json), so the graph itself carries your verdicts - and when you re-run, it commits them to a review manifest.

The routing rules are sensible defaults you can override: only-anomalous-silence lines default to "review," text/clipping/timecode failures default to "redo," and your manual choices always win.

Why the separation is the feature

Two outputs drive everything downstream: reviewed_batch (with all records attached) and approved_batch (only what passed). approved_batch is the one that flows to delivery - timeline rendering, export, packaging. Lines you didn't approve are set to review_hold in the reviewed batch, so no timeline or export can accidentally mix in unreviewed material. That's a deliberate safety: the README is explicit that unapproved items never silently enter a render. You also get retry_line_ids (feed to BatchRetry) and review_line_ids (the ones needing a human's second look).

The preview_limit (40) caps how many rows the table loads, mostly a UI convenience for very large batches.

Install

Pack install:

cd ComfyUI\custom_nodes
git clone https://github.com/T8mars/comfyui-fireredaudio-T8.git
cd comfyui-fireredaudio-T8
python scripts\setup_runtime.py

The node itself needs no model - listening and judging is file + UI work. QA upstream (if you want it pre-populated) needs the full model for ASR.

Gotchas

The two-run rhythm confuses people once. The table edits its serializable inputs on the front end, so your decisions aren't real until the node re-executes and writes the manifest. Click decisions, then run, then use the outputs. Also: when a repaired line comes back through, the old verdict moves to previous_human_review and the new take is judged fresh - the pack won't let a stale "approved" leak onto a re-generated line. And remember the honest limit: this gate checks QA + your ear, but it doesn't re-check loudness or format - that's DeliveryPreset's job on the way out.