Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
2
0
4
2
Current TTS evaluation metrics can be misleading for low-resource languages, as systems might "succeed" by speaking a neighboring language or producing unnatural audio that still yields a low WER.
Naive fine-tuning of Whisper beats more parameter-efficient methods by a mile for low-resource Pashto ASR, highlighting the importance of full model adaptation when pre-training is lacking.