AI Lab Pvt Ltd

Model · Speech recognition

sravaani-nepali-english-codemix-asr

AI Lab Pvt Ltd

Released by AI Lab Pvt. Ltd., Nepal. One of three models fine-tuned on the same Nepali–English code-mixed speech data: The output of every model on the test set (text only) is in Nepali speakers mix a lot of English into their speech. This model writes what they say the way people type it: Nepali in Devanagari, English words in Latin script. SraVaani-1.0 (449M parameters, hybrid TDT–CTC FastConformer) with all parameters fine-tuned. The tokenizer is unchanged. Best overall model of this release. - Nepali in Devanagari, English words in lowercase Latin script - no punctuation - numbers and symbols spelled out as words, in the language they were spoken in (ten percent or दस प्रतिशत) Held-out…

Dataset · Speech recognition

nepali-english-codemix-asr-test-outputs

AI Lab Pvt Ltd

Released by AI Lab Pvt. Ltd., Nepal. Text only: no audio is included. For every clip of our Nepali–English code-mixed test set (4,610 clips, 9.1 h) and of the FLEURS Nepali test set (586 digit-free clips), this dataset gives the reference transcript and the output of 7 speech recognition systems: SraVaani-1.0 zero-shot and fine-tuned for 6 and 12 epochs, and Whisper large-v3-turbo and IndicConformer, zero-shot and fine-tuned. Use it to inspect errors, compare models clip by clip, or recompute any metric. WER is computed over the whole set (total word errors divided by total reference words) on the normalised text. FLEURS writes English loanwords in Devanagari while our models write them in…