🌊 Audio Wave Analyzer

If you've ever wanted to change one word in a spoken clip without regenerating the whole thing, you need to tell the model exactly where that word sits in the audio - the start and end times, down to the millisecond. Eyeballing a waveform is miserable. The Audio Wave Analyzer does it for you: it chops an audio clip into timed regions (by silence, energy, or peaks) and hands you the timing data in the format F5-TTS's speech editor wants.

This is the timing front-end for F5-TTS editing workflows. F5-TTS can do targeted word/speech editing, but only if you can point at the segment. The analyzer is how you point.

How it works

You give it audio and pick an analysis method. It scans the waveform and detects region boundaries - for example, gaps of silence become the cut points between phrases. It draws an interactive waveform (that's what visualization_points controls: more points, more detail, heavier UI) so you can see and refine the regions, then it exports the boundaries as timing data. For fine work you can switch to manual and type the regions yourself.

The inputs and outputs that matter

• audio (AUDIO, optional) or audio_file (STRING path) - the clip to analyze. Wire in an AUDIO connection, or point at a file path.
• analysis_method - silence, energy, peaks, or manual. silence is the sensible default for speech (it cuts on the quiet bits between phrases). manual lets you define regions by hand via the fields below.
• precision_level - seconds, milliseconds, or samples. Milliseconds is the default and the right call for speech editing.
• export_format - f5tts, json, or csv. Leave it on f5tts if you're feeding an F5-TTS editing workflow; the other two are for when you want the timings elsewhere.
• options (ADV_AUDIO_OPTIONS, optional) - plug in an Audio Analyzer Options node here to tune the detection thresholds (how quiet counts as silence, how sensitive the peak finder is). Skip it and sane defaults apply.
• manual_regions / region_labels (optional) - for manual method, the region boundaries and their names.

Outputs: timing_data (STRING) is the one you'll use most - it feeds the F5-TTS editor. You also get processed_audio and segmented_audio (AUDIO) and analysis_info (STRING) for inspection.

How to install it

Ships with TTS Audio Suite. ComfyUI Manager: search 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

then restart. On Linux, install portaudio19-dev and libsamplerate0-dev first - the suite leans on libsamplerate for resampling audio, and skipping it is a common source of install pain. This node itself doesn't download a model; it's pure signal processing.

Common issues & troubleshooting

The regions are all wrong - too many, too few. That's a threshold problem, not a bug. Add an Audio Analyzer Options node and adjust: raise silence_threshold if quiet speech is being cut mid-phrase, lower it if whole phrases are being merged. energy or peaks methods can behave better than silence on noisy or music-bed audio.

The waveform UI is sluggish. Drop visualization_points. 2000 is the default; on a long clip you don't need 10000 to place a cut.

It cut on breaths and clicks. Speech is full of tiny sounds. Tune silence_min_duration (in the Options node) up so only real pauses count as boundaries, and clean obvious noise out of the clip first.