Just a few months in the past, we added non-public information to the Open ASR Leaderboard: 11 splits from Appen and DataOceanAI overlaying scripted and conversational speech from 5 accents (weblog). The neighborhood response was nice, with many highlighting the significance of held-out units for reliable evaluations. Since then, we added yet one more non-public cut up for Indian-accented English with Voice Enviornment (weblog).
We’re excited to announce a brand new collaboration with Liva AI so as to add much more non-public information with two new accents: Egyptian 🇪🇬 and Nigerian 🇳🇬.
| Dataset | Accent | Period [h] | Type | Transcription |
|---|---|---|---|---|
| Liva AI Conversational EG | Egyptian | 1.34 | Conversational, spontaneous | Punctuated, cased, with disfluencies. |
| Liva AI Conversational NG | Nigerian | 1.55 | Conversational, spontaneous | Punctuated, cased, with disfluencies. |
That is a part of our continued effort to enhance illustration in ASR evals (extra accents and extra languages) and to enhance belief (new non-public information to reduce benchmaxxing, and new information suppliers). We’re excited to see the neighborhood persevering with to push ASR efficiency for the world 🤗
Richer accent breakdown within the Non-public tab
With these new accents, we will now current a richer breakdown of accents, primarily based on Kachru’s Three Circles (kudos to Hume AI for this concept of their leaderboard). The Open ASR Leaderboard will now report:
- US vs. non-US (UK, Australia, Canada, India, Nigeria, Egypt)
- Internal Circle (US, UK, Australia, Canada) vs. Outer & Increasing Circles (India, Nigeria, Egypt)
As extra accents be part of, we’ll cut up the Outer and Increasing Circles.
About Liva AI
Liva AI collects multimodal interplay information by shopper merchandise utilized by actual folks. As a result of Liva owns each the gathering atmosphere and distribution, it might design merchandise round particular mannequin gaps throughout domains, industries, languages, modalities, and interplay sorts. This provides it entry to pure, high-context conduct that emerges by actual product use, quite than artificially elicited or task-based interactions.
