Lightweight speech recognition models are critical for edge deployment, yet highly optimized architectures like Moonshine often fail on morphologically rich, non-Latin languages such as Bengali. This study identifies the root cause of this failure as the model's English-centric byte-level tokenizer, which fragments Bengali words into high-fertility byte chains and triggers catastrophic autoregressive collapse during inference. To resolve this, a novel vocabulary transplantation pipeline is proposed to replace the decoder vocabulary with the native-script BanglaBERT WordPiece vocabulary and resize the corresponding token embedding matrix. Experimental results demonstrate a reduction in token fertility from 9.16 to 1.30. By decreasing autoregressive sequence length by 85.8%, decoding instability is entirely mitigated. When evaluated on the 882-hour Lipi-Ghor dataset, the modified architecture achieves a competitive 21.54% Word Error Rate (WER) and a Real-Time Factor (RTF) of 0.0053. Ultimately, this research provides a scalable, reproducible blueprint for cross-script adaptation of compact ASR models without the need for resource-intensive pre-training.
Kawshik Kumar Paul, Md. Nafiul Alam Fujics.CL cs.SD eess.AS
Speech systems used in customer-facing applications often require domain-specific language coverage. We present a synthetic Bengali speech dataset for telecom customer-care scenarios. The dataset contains 10,000 audio-text pairs, approximately 26.82 hours of 24 kHz speech, and predefined train, validation, and test splits of 9,000, 500, and 500 examples. It is publicly released on Hugging Face under the CC-BY-4.0 license. The speech was generated with OmniVoice in voice-cloning mode using a real female reference recording and transcript, with bfloat16 precision, 16 diffusion sampling steps, and a speaking-rate control value of 1.0. Along with the original Bengali text, the dataset provides a normalized transcript field designed for ASR/STT training and evaluation. We report an automatic intelligibility check over all 10,000 samples using a domain-adapted Whisper ASR model fine-tuned from bengaliAI/tugstugi_bengaliai-regional-asr_whisper-medium, along with a manual listening check on selected samples. The evaluation gives an average WER of 2.54%, an average CER of 0.59%, and median WER and CER values of 0.00%. These results suggest strong text-audio consistency under the selected automatic evaluation pipeline, while the paper also discusses the limitations of synthetic speech and STT-based evaluation.