Automatic Speech Recognition on In-house Hindi Conversational Speech (test)
14.6Agent WERIn-house SE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| In-house SECategory=Finetuning with Pseudo Labels +, No. of Params=94.4M2026.06 | 14.6 | 20.4 | 16.6 | |
| In-house SECategory=Speech Encoder finetuning, No. of Params=94.4M2026.06 | 16.2 | 24.2 | 19 | |
| NeMoCategory=Finetuning with Pseudo Labels +, No. of Params=114M/121M*2026.06 | 17.4 | 23.8 | 19.7 | |
| NeMoCategory=ASR finetuning, No. of Params=114M/121M*2026.06 | 19.6 | 29.9 | 23.3 | |
| SpeechBrain ModelCategory=Trained from Scratch, No. of Params=94.4M2026.06 | 22.9 | 39.7 | 28.9 | |
| MMSCategory=Speech Encoder finetuning, No. of Params=1B2026.06 | 27.4 | 34.6 | 29.4 | |
| MMSCategory=Speech Encoder finetuning, No. of Params=300M2026.06 | 28.6 | 38.9 | 31.5 | |
| Google TelephonyCategory=Commercial, No. of Params=-2026.06 | 34.3 | 39.9 | 36.9 | |
| Data2Vec AQCCategory=SoTA Open Source, No. of Params=100M2026.06 | 40.9 | 48 | 43.7 | |
| Whisper-large-v3Category=SoTA Open Source, No. of Params=1.5B2026.06 | 47.2 | 53.2 | 49.2 | |
| NeMoCategory=SoTA Open Source, No. of Params=114M/121M*2026.06 | 49.1 | 41.8 | 44.4 |