Tokenizer choice is crucial for non-Latin languages in edge ASR—swapping an English byte-level tokenizer for a native-script WordPiece vocabulary can eliminate autoregressive collapse and enable efficient multilingual deployment without retraining.
This paper fixes a critical problem where lightweight speech recognition models fail on Bengali by replacing the English-focused tokenizer with a Bengali-native one. The fix reduces token fragmentation from 9.16 to 1.30 tokens per word, eliminates decoding instability, and achieves competitive accuracy (21.54% WER) while maintaining fast inference (RTF 0.0053) on edge devices.