Deepgram expands Nova-3 with 10 languages and keyterm prompting
The verbal AI adds 10 monolingual languages to its latest generation, Nova-3, and enables multilingual keyterm prompting to improve recognition.
Deepgram, the verbal AI specialist, has added support for 10 monolingual languages to its Nova-3 automatic speech recognition model, and also launched multilingual Keyterm Prompting.
The update targets languages with tonal variation, complex morphology and multi-script writing systems while preserving low latency across both batch and streaming modes, Deepgram explained in a statement.
These newly added languages span Southern and Eastern Europe, the Baltics and Southeast Asia. They include Greek (el), Romanian (ro), Slovak (sk), Catalan (ca), Lithuanian (lt), Latvian (lv), Estonian (et), Flemish (nl-BE), Swiss German (de-CH) and Malay (ms).
Deepgram said Nova-3 reduces word error rate versus its predecessor, Nova-2, in these languages, with streaming transcription often showing the largest relative gains. Malay, Romanian and Slovak recorded some of the biggest reductions, in several cases exceeding 20%, it added. And, no retraining is required, according to the company.
Nova-3 Multilingual accepts up to 500 tokens (about 100 words) of key terms in a single request and can prioritise those terms across all supported languages, it noted.
Deepgram is offering $200 in credits for new sign-ups to Nova-3, which is enough to cover over 750 hours of transcription or 200 hours of speech-to-text.
The recap
- Deepgram adds support for 10 monolingual languages to Nova-3.
- Multilingual Keyterm Prompting accepts up to 500 tokens per request.
- Users switch languages via API using the listed language codes.
