Controllable multilingual speech recognition with verbatim or clean output and word-level timestamps, from Nyra Labs.
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.
CrisperWhisper 2.0 is a Whisper-based speech recognition model from Nyra Labs (nyra health GmbH). It succeeds CrisperWhisper 1.0, which supported only English and German and produced verbatim output only. Version 2.0 adds an explicit switch between a verbatim transcript (fillers, repetitions, stutters, false starts, vocal sounds such as laughter) and an "intended" transcript that drops disfluencies and formats numbers, dates and emails. It covers most languages Whisper supports.
Word-level timing is obtained by supervising selected cross-attention heads with ground-truth word timings. The vendor reports 29.6 ms mean boundary error on TIMIT and 40.6 ms on Buckeye, lower than the commercial systems it compared against (ElevenLabs Scribe v2, Deepgram Nova-3, xAI, Cartesia). Longform audio is handled with "conditional continuation", where the last few words are fed back to the model as the window advances. A "Verbatimize" approach is described for adding verbatim annotations to existing speech corpora. Nyra also published an evaluation suite, the nyra verbatim speech benchmark, scoring disfluency F1 across ten languages. A "Pro" variant is reported as ranking first in every tested language; these results are vendor-reported.
Access is through a pay-as-you-go hosted API with a prepaid wallet, priced by audio length (up to one hour per request), and a free browser demo needing no account. Each call returns both the verbatim and intended transcripts with timestamps and alignment. A YouTube review describes an open-source release in sizes from 0.2B to 2B parameters, but no weights or repository links appear in the provided pages, so open-weights status is unconfirmed. Self-hosting and on-device use are offered via custom plans.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open-source transcription model with verbatim and intended modes plus word-level timestamps. Creator tested it on a Hugging Face space; it beats ElevenLabs on its own benchmark. Four sizes from 0.2B to 2B parameters.
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 · Edge0
Open-weights 0.17B-parameter multilingual text-to-speech model with zero-shot voice cloning, runs on CPU via ONNX.
Model · IndexTeam
Open-weights zero-shot text-to-speech model from Bilibili's IndexTeam with voice cloning and emotion control in 5 languages.
Model · Cactus
Compact 16.9 MB speech-to-text model from Cactus that runs on CPU with no dependencies.
Model · Google
Google's speech-to-text model with verbatim and smart modes, speaker detection, word timestamps and a live streaming variant.
Model · Microsoft
Microsoft AI's second-generation speech-to-text model for transcribing noisy, domain-specific audio.
