01Cactus Compute · Oct 2Cactus Whistle: 16.9 MB Apache-2.0 ASR, tested against Whisper base
Whistle is a speech-to-text model released on Hugging Face (repo created Sep 30) with a launch blog on Oct 2. License on the card: apache-2.0. It is a single 16.9 MB file (whistle.cact) that runs on CPU through Cactus's Needle engine. Access is open download, no gate. It transcribes 16 kHz mono audio up to 30 s per pass in English, German, French, Spanish, Italian, Dutch and Polish, and also returns word timestamps and encoder embeddings. Parameter count is not stated on the card or blog.
Benchmarks are the Cactus team's own claims. Whistle's WER was measured over 86,174 utterances with Whisper normalizers; the Whisper and Moonshine figures are published numbers from the multilingual Whisper base checkpoint, not large-v3. The blog says Whistle is ahead of Whisper base on LibriSpeech test-clean/test-other, SPGISpeech, Earnings-22 and the FLEURS average, and behind on TED-LIUM, AMI and the MLS average. Comparison against your Whisper large-v3 is not available. The card says no test audio appears in the training or validation data. Training data sources are not named in the pages I read.
For your stack: worth a test only if you need edge or embedded ASR; no evidence here that it beats Whisper large-v3 on accuracy.
02FermionResearch · Sep 28Phonon-2 claims 5.21% average WER at 164 MB, CC-BY-4.0
Phonon-2 (FermionResearch/Phonon-2, repo created Sep 28, outside the seven-day window but within 30 days) is a low-bit quantized derivative of NVIDIA parakeet-tdt-0.6b-v3 for English, in MLX format. License on the card: CC-BY-4.0 (same as the base model); the CLI and repo code are Apache-2.0. Download is 164 MB, ungated.
Self-reported on the Open ASR Leaderboard's seven English sets (LibriSpeech clean/other, AMI, Earnings-22, GigaSpeech, SPGISpeech, VoxPopuli): 5.21% average WER, with LibriSpeech clean 1.72 and other 3.92. The card's table includes Whisper large-v3-turbo at 6.58 average (a leaderboard row), not large-v3, so a direct large-v3 comparison is not given. The card says the uncompressed teacher scores 4.96. Note the model is English-only, so it does not cover the EU-language side of your requirements.
For your stack: worth testing as a small English alternative to Whisper large-v3 on your own audio; the benchmark is the authors' claim, not independently reproduced.
03Poseidon AI · Oct 2Numo Indic Speech: 3,845.5 hours, CC-BY-4.0, but gated and not EU data
psdn-ai/numo-indic-speech (v1.0, created Oct 2) is a crowdsourced, scripted read-speech corpus: 353,349 recordings, 3,845.5 hours across Bengali (745.1 h), Hindi (1,314.6 h), Tamil (561.6 h) and Telugu (1,224.2 h), with transcripts, pseudonymous speaker IDs, age/gender/country metadata, per-recording WER/CER and a synthetic-speech suspicion score. Audio is embedded WAV in Parquet; total size 1.23 TB.
License per the card: CC-BY-4.0, commercial use allowed with credit. Access is auto-gated: you must accept sharing contact details, and the card says Poseidon may contact you about related datasets. The data is Indic-language, so it does not satisfy an EU-data requirement. The card names no scraped sources; it describes crowdsourced collection but I did not find a consent statement. It reports no ASR benchmark results.
For your stack: skip for EU-language work; only relevant if you extend to Indic languages, and check the gate terms first.