FireRedAudio 语音编辑 · T8star-Aix

Speech isn't audio you edit with scissors once a model has generated it - but FireRedAudio can rewrite it. SpeechEdit gives you two ways in: semantic editing, where you describe the change in plain language ("remove the cough", "change '今天' to '明天'", "replace that sentence"), and acoustic editing, which applies the official template parameters for pitch, speed, and volume. It's the node that turns a finished take into a fixed take without regenerating the whole line from scratch.

How it works

With edit_type set to semantic (default), your instruction string goes to the model along with the audio, and the model regenerates the edited region - an insert, delete, or replace expressed as natural language. With acoustic, the same node applies the model's built-in acoustic controls instead; you still pass an instruction, but it's parameterized (the official template's pitch/speed/volume adjustments) rather than free-form. Either way the output is a new WAV written by the worker, plus the edited text and a run report.

The inputs:

• model - required; needs the full model (RedAE decoder) for editing.
• audio - the clip to edit.
• instruction - the edit command. This is where the quality lives: "delete the second sentence" is a much better instruction than "make it better."
• edit_type - semantic or acoustic.
• settings - optional generation parameters from the pack's 生成参数 node.

Outputs: 编辑后音频 (AUDIO), 语义编辑文本 (the resulting text, mainly meaningful for semantic edits), and 运行报告.

Where it fits

The pack's local-repair loop is the intended home: 局部修复范围 (pick a range by seconds or from a locator JSON) → this node (edit it) → 局部修复回填 (crossfade it back in, non-destructively, with source SHA-256 verified). The README's 21_podcast_local_repair example runs exactly that - a hand-picked range, a semantic fix, an equal-power crossfade back into the original file that's never overwritten. If you have a single blemished line that QA flagged, fixing it in place with this node beats re-rolling the dice on a whole new generation - and it preserves the source channels and sample rate when done through the repair loop.

Installing and the honest caveat

Pack-wide install (ComfyUI Manager search comfyui-fireredaudio-T8, or clone + python scripts\setup_runtime.py), full model profile. The caveat is expectation management: semantic speech editing is the model's suggestion of what the region should sound like, not a surgical cut. For a clean "remove the cough in the first line" it's excellent; for "reperform this sentence with different emotion" you're asking the model to act, and the result can drift from the original voice's delivery. When the edit changes the performance too much, that's when you regenerate the line with SeedAudition instead. Verify the edited text output - it tells you what the model thinks it did, and it's usually right.