Login
You're viewing the pc.cafe public feed.
  • Nov 16, 2026, 12:14 PM

    When I launched Cava Reader I said it was my first closed-source project. The app still is, but the part that does the actual cool stuff isn't anymore :blobcat:

    Meet CavaAlign v1. You give it a recording and its text, and it tells you when every single word is said. It's tiny, it's fast, and it's the model I made for read along in Cava Reader, and it's now open source! (under AGPL-3.0)

    One 16 MB file, 35 languages, it can run on a phone's CPU. Numbers are in the image attached, and yes, that's a 16 MB model next to a 378 MB and a 967 MB one, doing better than them, I double-checked lol

    It's made for books, so footnotes and the "this is a LibriVox recording..." intros don't throw it off, it’s able to just skip them.

    The old engine in Cava Reader needed a 40 MB download for every extra language except English (which was bundled). This one is 16 MB for all 35, and it ships inside the app binary, the size of the whole app is a smaller size too

    The model weights, a Python example aligner, a Kotlin port for Android, and a guide for mobile apps are all on Hugging Face so it’s easy to get started using it in your own projects:
    huggingface.co/cava-reader/cav

    It’s a standard ONNX model, so it runs anywhere the ONNX Runtime does: phones, desktops, servers, Raspberry Pis, and it should even run in a browser. No GPU needed, it's fully offline, and it handles recordings of any length with flat memory use. If you have audio and its text, it'll time every word.

    if you build something with it, please show me!! <3

    #OpenSource #MachineLearning #Accessibility #Audiobooks #AGPL #CavaAlign

    CavaAlign v1: word-level timing for any recording and its text. 16 MB, one file for 35 languages. 88.0% of English word starts within 50 ms, 0.3% more than 250 ms off or untimed. 86.7% within 50 ms averaged over all 35 languages.

Compared with other models (size, languages, within 50 ms, more than 250 ms off or untimed):
CavaAlign v1: 16 MB, 35 languages, 88.0%, 0.3%
wav2vec2-base (WhisperX): 378 MB, English, 84.4%, 0.4%
PocketSphinx: 10 MB, English, 83.8%, 8.6%
Vosk small-en: 71 MB, English, 80.4%, 12.7%
MMS-300m: 1.26 GB, 1,100+ languages, 77.0%, 7.2%
NeMo Conformer-CTC S: 46 MB, English, 67.3%, 5.7%
Whisper small.en: 967 MB, English, 48.3%, 10.8%

Benchmark: 70 LibriSpeech speakers, 12,195 English words, exact text given to every system, scored against Montreal Forced Aligner timings. Two independent reference alignments agree on 89.5% of words.
    💬 1🔄 12⭐ 13

Replies

  • Nov 16, 2026, 12:16 PM

    For the ML nerds, what's inside:

    A raw-waveform front end into an 8-layer streaming conformer (d 256, 14.6M params), with a CTC head over just a-z and an apostrophe. Text gets romanized per word first (Cyrillic, umlauts and accents all become a-z), which lets the tiny vocabulary cover all 35 supported languages with nothing extra needed per-language.

    int8 weights, standard ONNX ops, plain ONNX Runtime on CPU. (CPU chosen for compatibility on older devices)

    On top of it sits a monotonic dynamic program with text-skip, audio-skip and front-matter arcs. That's the part that lets it deal with footnotes, chapter titles and audio that doesn't quite match the text with no problems.

    About the benchmark: it's real audiobook speech (70 LibriSpeech speakers) with the exact text, the reference timings come from Montreal Forced Aligner, not from someone marking every word by hand. Two independent MFA runs on the same audio only agree on 89.5% of words (the dotted line in the image), so at around 90% and up you can't really tell how much better something is anymore. How it does with messier text is on the model card too.

    Just to be clear it’s only an aligner, it doesn't transcribe. It’s for Latin and Cyrillic scripts only. And it's made for read speech like audiobooks, not conversations.

    Per-language numbers and the full benchmark method are on the model card on Hugging Face.

    💬 0🔄 1⭐ 2