PitchIn: Natural-Language Singing Coaching
from Unstructured Community Feedback

Audio examplesTry the demo
PitchIn overview: community singing feedback is reconstructed into Focused QA, Overall Feedback, and Recognition QA. Audio–language models are adapted with a frozen audio encoder and language-model LoRA.
PeerCoach construction and PitchIn training. PitchIn-Q and PitchIn-M use FQA + OF; Recognition QA is an additional experimental condition.

PitchIn learns to give natural-language singing feedback from observations and advice shared in singing communities. We build PeerCoach by reconstructing these interactions into instruction examples, then adapt Qwen2.5-Omni and MOSS-Music using Focused QA and Overall Feedback. At inference time, the model needs only a recording and a question.

Audio Examples

Selected full recordings with unedited responses from pretrained Qwen2.5-Omni-7B and PitchIn-Q. Notes compare the responses with expert references.

Question

Pretrained

PitchIn-Q

FQA + OF

Try PitchIn

Upload a singing clip and ask a question. English questions recommended.

Connecting to the demo…

Open on Hugging Face ↗

BibTeX
Manuscript, 2026