Ursa 2
Ursa 2 was announced on 11 October 2024. The release enlarged the self-supervised model and added Irish and Maltese support.
Access not confirmed
Pretraining followed by language training
The release describes more unlabelled pretraining followed by targeted training for specific languages. It also moved inference fully to GPUs. The body-model parameter count does not specify the size of every component in the recognition system.Speechmatics: Ursa 2, 11 October 2024
Practical guidance. Ursa 2 is useful history for understanding Standard and Enhanced. An old evaluation of this generation should not be presented as a measurement of the continually updated API today.Speechmatics: Ursa 2, 11 October 2024
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Access not confirmed |
| Launch date | |
| Model identifiers | Ursa 2 models |
| Self-supervised body model | 2.88 billion parameters; trained with 2.8 million audio hoursSpeechmatics: Ursa 2, 11 October 2024 |
| Broad service availability | 14 October 2024 announcement says Ursa 2 powered all speech servicesSpeechmatics: Cascaded conversational AI and Ursa 2 availability |
| Exact checkpoint access | No current frozen Ursa 2 checkpoint verifiedSpeechmatics documentation: Models |
Related models
- Ursa
Speechmatics · Ursa
Historical recognition generation with self-supervised audio representations and separate decoding components.
Launch date Access not confirmed - Enhanced
Speechmatics · Enhanced
General speech recognition model that prioritizes accuracy across supported language packs.
Launch date not confirmedAPI · Provider product - Standard
Speechmatics · Standard
General speech recognition model that prioritizes transcription throughput.
Launch date not confirmedAPI · Provider product