Speech recognition model trained for noisy, reverberant and distorted real-world audio, with RL post-training.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Mega-ASR is an automatic speech recognition model aimed at difficult acoustic conditions. It is trained on data covering seven atomic acoustic effects (reverberation, echo, additive noise, far-field, frequency dropout, bandwidth limitation, clipping) and 54 compound environmental scenarios built from them.
Training uses a two-stage recipe: Acoustic-to-Semantic Progressive SFT (A2S-SFT), followed by Dual-Granularity WER-Gated Policy Optimization (DG-WGPO), a reinforcement-learning stage combining a token-level refinement reward with a sentence-level reconstruction reward. The training set, Voices-in-the-wild-2M, is synthesized via spectral manipulation with an agentic plausibility check, and samples above 70% WER are filtered out.
The project page shows side-by-side examples against Qwen3-ASR, Gemini-3-Pro, Seed-ASR and Whisper on empty-output recovery, long utterances, babble noise, restaurant noise, financial entities and short phrases. These are curated demos, not full benchmark results. A YouTube review mentions a size under 5 GB and fine-tuning scripts, and a claim of nearly 30% gains in hard conditions; the provided pages do not confirm weights, license or code release details.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Speech recognition model for noisy real-world audio, trained on 2.6M samples across seven acoustic problems. Claims nearly 30% gains over leading models in difficult conditions; under 5 GB with fine-tuning scripts.
Closest neighbours by meaning, not just by category.
Model
Open-source append-only streaming speech recognition model with configurable low-latency decoding, English and Mandarin.
Model · Fermion Research
Open 164 MB English speech recognition model from Fermion Research, derived from Parakeet TDT 0.6B v3.
Model · Microsoft
Microsoft AI's second-generation speech-to-text model for transcribing noisy, domain-specific audio.
Model · Nyra Labs
Controllable multilingual speech recognition with verbatim or clean output and word-level timestamps, from Nyra Labs.
Model · Microsoft
Microsoft AI's streaming speech-to-text model for near real-time transcription, succeeded by MAI Transcribe 2.
Model · Cactus
Compact 16.9 MB speech-to-text model from Cactus that runs on CPU with no dependencies.
