crisp-models

rafay99-epic/crisp-models

crisp-models is a audio classification model from rafay99-epic. Use it for the audio classification task on the model card, and read the license before you ship it in a product. The card lists the license as cc-by-4.0.

Wren is the fast, lightweight model in Crisp's filler-detection family (named after birds — keen ears, clean song; the larger, higher-accuracy sibling is Kestrel). A wren is the tiniest bird yet punches far above its…

Downloads · 30 days

8% of all-time downloads

All-time downloads

Public

Repo size

2.4 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.pt545 KB · 51%

At a glance

Task: Audio Classification
License: cc-by-4.0
Access: Public
Created: Jun 21, 2026
Updated: Jun 25, 2026
SHA: efd786b0

Model card

From the Hugging Face model README

Wren — Crisp filler detector

Wren is the fast, lightweight model in Crisp's filler-detection family (named after birds — keen ears, clean song; the larger, higher-accuracy sibling is Kestrel). A wren is the tiniest bird yet punches far above its size — fitting for a ~75k-param CNN that hits 0.94 precision at ~600× real-time.

It detects filler words ("uh", "um") in speech for the Crisp macOS app — a fast, on-device alternative to running full speech-to-text just to find fillers.

Input: one log-mel chunk, [1, 1, 64, 25] (250 ms of 16 kHz mono audio, standardized with fixed constants MEL_MEAN=-18.5658, MEL_STD=17.9252).
Output: filler_prob — P(this chunk is a filler), 0…1.
Format: a single-file Core ML .mlmodel (download directly, no unzip). Runs on-device — Core ML schedules it across the Neural Engine, GPU, or CPU.

Feature extraction (mel + chunking + the slide/merge into time ranges) lives in the host app; the model itself is a pure chunk → probability function.

Training

Data: PodcastFillers — ~35k human-labeled "uh"/"um" events, 350+ speakers. Trained on the train split.
Task: binary filler-vs-speech on 250 ms chunks (silence/pauses are handled separately by the app via ffmpeg silencedetect).

Evaluation (held-out, episode-disjoint `test` split)

Threshold	Precision	Recall	F1
0.5	0.914	0.933	0.924
0.7	0.944	0.894	0.918
0.8	0.959	0.863	0.908

Speed: ~600× real-time on CPU. Recommended operating threshold 0.7 (Crisp favors precision — a false positive cuts real speech). Main error mode: occasional confusion of short real words; breaths/laughter/music are rarely misfired.

Limitations

English only. Trained on podcast audio; very different domains (heavy accents, noisy rooms) may need a higher threshold or a fine-tune. Not a transcriber — it only flags fillers.

Files & versions (open weights)

Everything is downloadable — pick whichever you need:

File	What
`Wren.mlmodel`	Core ML model (Apple Neural Engine). A single file the app downloads directly.
`Wren.pt`	Raw PyTorch weights (`state_dict`) — for your own inference/fine-tuning.
`Wren.config.json`	Audio framing + normalization + input/output names + recommended threshold.

Versioning: the version is the repo's commit count, v0.0.N (mirrors Crisp's 0.<commits> scheme); each release is one commit + one tag. Pin an exact version in the URL:

https://huggingface.co/rafay99-epic/crisp-models/resolve/v0.0.5/Wren.mlmodel

main always points at the latest.

License

Code: GPL-3.0 (Crisp). Model weights derive from PodcastFillers (CC). Credited to Syntax Lab Technology / Abdul Rafay.

Task: Audio Classification
License: cc-by-4.0
Languages: en
Created: Jun 21, 2026
Updated: Jun 25, 2026