Real-time multilingual ASR using rolling buffers and monolingual models [P]
Researchers have developed a novel approach for real-time multilingual Automatic Speech Recognition (ASR) that utilizes rolling buffers and specialized monolingual models. Instead of a single, large multilingual model, this system routes audio to smaller, efficient monolingual models (~100M parameters each) for transcription. This method achieves a Word Error Rate (WER) of approximately 13% on inter-utterance code-switching benchmarks, outperforming tested cloud APIs and other systems. AI
![Real-time multilingual ASR using rolling buffers and monolingual models [P]](https://preview.redd.it/qu5jir6i0p4h1.png?width=140&height=20&auto=webp&s=94ba13db0fc9733c3f14568ff194b6e737667a72)
IMPACT This approach offers a more efficient and accurate solution for real-time multilingual speech recognition, potentially improving accessibility and usability of voice-enabled applications across different languages.