#speech-recognition
Every summary, chronological. Filter by category, tag, or source from the rail.
Data Scale, Not Latency, Drives Cross-Lingual ASR Transfer
Multilingual encoder initialization provides a significant performance boost for streaming ASR only in low-data regimes; as target-language data scales, the advantage of multilingual over English-only initialization vanishes, regardless of latency constraints.
Microsoft's MAI-Transcribe-1.5: Production-Ready Speech Recognition
Microsoft's MAI-Transcribe-1.5 improves speech-to-text with 43-language support, 5x faster long-form inference, and entity-aware keyword biasing for enterprise accuracy.
NVIDIA's Nemotron 3.5 ASR: Efficient Multilingual Streaming Speech
NVIDIA's Nemotron 3.5 ASR is a 600M-parameter, cache-aware streaming model that transcribes 40 languages in real-time from a single checkpoint, offering configurable latency-accuracy trade-offs without retraining.
Building Robust Voice AI: Beyond Simple Transcription
Speaker diarization is essential for understanding conversations, but combining it with transcription is difficult due to overlapping speech, mismatched timestamps, and poor generalization of ASR models to multi-speaker environments.
AI EngineerShowing 4 of 4