Pitch Feature Extractor ⚡🅡🅞🅣🅘

AudioFeatureExtractor covers loudness, brightness, and rhythm-adjacent stuff; this node is specifically about pitch - which note, or how high a frequency, is present in the audio at each moment. If you want an effect that follows a vocal melody, gets more intense as a bassline drops lower, or catches a singer's vibrato, this is the extractor built for it.

The five extraction methods

extraction_method is where the personality lives: frequency gives you the raw pitch in Hz, best for closely following a melody. semitone rounds that to the nearest musical note, which behaves more predictably for musically-timed effects than a continuous Hz value would. pitch_direction tells you whether pitch is rising or falling, which is a great driver for directional animations - something that visibly moves up as a note climbs. vibrato_signal and vibrato_strength detect the wobble singers and string players add on sustained notes - the signal itself for tremolo-style effects, or the strength value if you just want intensity, not the raw wobble.

Pitch detection under the hood

opt_crepe_model (none, tiny, small, medium, large, full) picks how the actual pitch tracking is done. CREPE is a well-known neural pitch tracker - picking anything other than none swaps in that model instead of basic pitch detection. It's a straightforward speed/accuracy trade: tiny/small run fast but track less precisely, medium is the balanced default, large/full are the most accurate and the slowest. Worth knowing: crepe/torchcrepe isn't in the pack's pinned requirements.txt - if switching opt_crepe_model away from none throws an import error, you may need to install that dependency into your ComfyUI environment yourself.

opt_pitch_range_collections (optional) restricts detection to specific frequency windows, built from this pack's PitchRange, PitchRangeByNoteNode, or PitchRangePreset nodes. Without one, the extractor listens across its whole detectable range; with one, it can ignore, say, everything outside a vocalist's actual register - useful for cutting out low-end rumble or high-frequency noise that isn't the pitch you actually care about.

The inputs and outputs that matter

• audio (required, AUDIO) - the source.
• extraction_method - frequency / semitone / pitch_direction / vibrato_signal / vibrato_strength.
• opt_crepe_model - pitch-tracking accuracy vs. speed.
• opt_pitch_range_collections (optional, PITCH_RANGE_COLLECTION) - narrow the detectable range.
• frame_count (default 0) - auto-calculates from audio length and frame_rate; set explicitly to interpolate to a target length instead.

Output is a single FEATURE - wire it into any Flex target's reactive input, FeatureToMask for a plain mask, or PitchVisualizer to burn the value onto video for a sanity check.

How to install it

ComfyUI Manager: search "RyanOnTheInside", install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ryanontheinside/ComfyUI_RyanOnTheInside
cd ComfyUI_RyanOnTheInside
pip install -r requirements.txt

then restart.

Common issues & troubleshooting

Feature is flat or noisy on a busy mix. Pitch detection assumes something close to a single dominant pitch at a time - polyphonic music (a full band, layered vocals) confuses any monophonic pitch tracker, CREPE included. Isolate the vocal or lead instrument first if you can (a stem separator, or a cleaner solo recording) for a usable signal.

opt_crepe_model set to large/full and processing crawls. That's the expected trade-off - the larger CREPE models are meaningfully slower. Drop to medium or small unless you specifically need the extra accuracy.

Detection misses notes outside where you expect. Check opt_pitch_range_collections isn't accidentally excluding the range your source actually sits in - a range built for a bass part will silently miss a soprano vocal layered on top.