The primary operational bottleneck stems directly from relying on standard, built-in device microphones. In an uninsulated room with ambient room noise, acoustic reflections, or sound leaking from open-backed headphones, the local neural network's pitch detection confidence score can drop. This tracking interference can trigger artificial accuracy penalties on the leaderboard, even if you hit the correct note.
Additionally, the automated source-separation pipeline faces clear technical boundaries when processing dense, multi-layered, or highly compressed audio files. If your uploaded track features heavy autotune modulations, dense wall-of-sound guitar tracking, or complex multi-part choral harmonies, the pitch extraction tool can generate erratic, jagged note nodes on the display that don't quite track the primary melody.
Finally, while the gamified approach is highly effective for building ear-to-voice accuracy, the interface cannot monitor or correct physical vocal mechanics. An operator can easily hit the exact target pitch coordinate while using improper breathing support, keeping their jaw locked, or over-singing with dangerous throat tension—actions that will still yield a high digital score but can cause physical fatigue or vocal cord strain over long practice sessions.
the on-device pitch detection makes sense for privacy on your own voice, but curious about the song side - when someone uploads a commercial track to extract the melody, does that audio get processed on a server to pull out the vocal line, or is the melody extraction also happening locally? seems like the copyrighted song itself is the part more likely to need to leave the phone, separate from your own singing
Being able to upload a song and have it figure out the key and break it into phrase-by-phrase practice is honestly such a smart idea, would have saved me so much time with my vocal coach. One thing that would make it even better is if you could slow down tricky sections without changing the pitch, similar to how some guitar tab apps let you loop a riff. That way you could actually nail the runs before speeding back up to the original tempo.
Would love a way to save multiple range tests over time and see the trend in a simple graph, super motivating to actually watch my range expand song by song. Also maybe tag which songs tripped me up so I can circle back to them.
Love the Guitar Hero angle for singers, this could really make practice stick. One thing I'd want is the ability to export a session as a MIDI or PDF practice sheet so I can review my pitch accuracy offline or share it with my vocal coach.
Played around with the free range test and the Guitar Hero-style feedback is genuinely satisfying, makes it way easier to see where I'm flat. Surprised how smooth the pitch detection felt in the browser without any setup.
Does the real-time pitch feedback actually work through the mic in a regular browser tab, or do I need to download something for low latency?