--- title: tonesearch colorFrom: indigo colorTo: purple thumbnail: https://raw.githubusercontent.com/aipdv/tonesearch/main/assets/banner.jpg sdk: gradio sdk_version: 6.30.0 python_version: "3.12" app_file: app.py license: apache-2.0 short_description: Search call audio by what was said and how it was said --- # tonesearch demo Search 8 call recordings by **what** was said (EmbeddingGemma 2, audio in, no transcript) and **how** it was said (emotion2vec+ tone score, prebuilt). - Search tab: type a phrase, pick a tone, set the tone weight. - Timeline tab: tone over one call, with the top 3 spikes. - Your call tab: upload your own call (up to 5 min). Both models run on a ZeroGPU GPU, then you get its tone timeline and top segments. Demo audio: 8 calls from "HaessigDB: A Database of Irritable Speech with Intensity Grading" by Niklas Weller, Marc Christopher Grau and Ivo Blohm (2026), doi:10.5281/zenodo.18745894, licensed CC-BY-4.0 (https://creativecommons.org/licenses/by/4.0/). The calls are acted, not real customers. The sentence clips were joined into whole calls. Models: `google/embeddinggemma-2` (query and audio embeddings), `emotion2vec/emotion2vec_plus_base` (tone scores; FunASR model license). Demo call scores are prebuilt; uploads run live.