iPhone app

self.mp3 1.0.0 (1874) for iPhone, from 90a47db. Installed from the install page, not from here: https://xiao215.github.io/selfmp3/install/

Android app

self.mp3 1.0.0 (1874) for Android, from 90a47db. Installed from the install page, not from here: https://xiao215.github.io/selfmp3/install/#android

Listening model (CLaMP 3), v1

The listening model's files for self.mp3's server: CLaMP 3 (with MERT under it), exported to ONNX and made 8-bit by scripts/sound-models/export.py. The server downloads them into its data folder the first time analysis wants them and checks each against the hash written in apps/server/src/sound/models.ts. Nothing here needs downloading by hand.

File	What it is	Size
mert.onnx	MERT-v1-95M: one 5 s chunk of 24 kHz audio → its vector (8-bit matrix multiplications, convolutions in full precision)	122 MB
clamp3_audio.onnx	CLaMP 3's audio encoder: a clip's chunks → the song's vector (8-bit)	87 MB
clamp3_text.onnx	CLaMP 3's text encoder (XLM-RoBERTa base): a description → its vector (8-bit vocabulary only)	535 MB
tokenizer.json, tokenizer_config.json	XLM-RoBERTa's tokenizer, as published	9 MB

Checked against CLaMP 3's own Python pipeline on 585 songs: these 8-bit files give the same search and "sounds like" results as the full-precision model.

Licences. CLaMP 3: MIT (sanderwood/clamp3). XLM-RoBERTa: MIT (FacebookAI/xlm-roberta-base). MERT-v1-95M: CC BY-NC 4.0 (m-a-p/MERT-v1-95M), so these files are for non-commercial use.