Speech-to-text models

66 models on Hugging Face (WindyWord org): 10 general tiers built on OpenAI Whisper and Distil-Whisper, 49 single-language models, and 7 whisper.cpp builds. None are benchmarked by this site yet.

Honest status: the tier models redistribute upstream Whisper / Distil-Whisper weights in more formats; the model cards say no fine-tuning was applied. The single-language "lingua" models were imported from community Whisper fine-tunes. Each model page names its upstream and that upstream's licence (read from its Hugging Face repository on 2026-10-01); 5 of them come from an upstream that declares no licence and are marked under review: do not use those commercially until the note is removed.

General tiers

Distil-Whisper bases are English-only. The three listen-windy-distil-* repos are MIT, matching their Distil-Whisper upstream (corrected 2026-10-01).

Single-language models

whisper.cpp (ggml) builds

Run one

pip install transformers torch huggingface_hub   # plus ffmpeg on your PATH to read audio files

from huggingface_hub import snapshot_download
from transformers import pipeline

# Weights live in the safetensors/ subfolder of the repository.
path = snapshot_download("WindyWord/listen-windy-nano", allow_patterns=["safetensors/*"])
asr = pipeline("automatic-speech-recognition", model=f"{path}/safetensors")
print(asr("audio.wav")["text"])

Pattern run on 2026-10-01 against WindyWord/listen-windy-nano with transformers 4.57.6 and 5.18.0 on CPU, on a 16 kHz mono WAV. CTranslate2-only repos (-ct2) use faster-whisper; their pages show how.