AI Briefing#speech-recognition
The latest AI and developer news about #speech-recognition, with the original source and a short summary.
The latest AI and developer news about #speech-recognition, with the original source and a short summary.
Microsoft launched MAI-Transcribe-2-Streaming, which ranks No. 1 on Artificial Analysis, along with two new voice models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash, designed for low-latency conversational AI.
The new method improves accuracy by 20.8% in-domain and 10.0% cross-domain over prior methods by coupling online teachers with server-side update stabilization.
The method achieves 2.8× compression while staying within 1.9% relative WER of the teacher model on five of six pairs.
An open-source project has been released that enables running NVIDIA's Parakeet speech recognition model on the JVM with native-like performance.
NVIDIA has released Nemotron 3 Diarization, a 100M parameter open-weight model that ranked first on the Voice Arena Diarization-Bench.
Wispr Advanced Interfaces Lab released 'Canto', a voice model that achieves low Word Error Rate (WER) in real-time transcription environments by applying GRPO reinforcement learning.
Orukeet, a 25-language ASR model based on NVIDIA Parakeet TDT, has been released, improving average WER by 10.6%.
Conversational AI combining LLMs and speech recognition is emerging to improve the low resolution rates of fixed-menu-based IVR.
The open-source tool hayamimi has been released, generating real-time multilingual subtitles using only the CPU without GPU dependencies.
Microsoft released its in-house developed MAI-Transcribe-2, offering a low price of 10 cents per hour and speeds 5–10 times faster than competitors.
Meta released the real-time streaming ASR model 'Muse Voice Transcribe', achieving a top benchmark ranking with a 3.1% WER.
Hugging Face has released Monsoon, a speech recognition evaluation dataset for Hindi and Indian English, to support Global South languages.
Speaker diarization is a technology that identifies who spoke when in an audio recording.
FireRedTeam released FireRedAudio, a 9B parameter unified audio language model that decouples understanding and generation.
A study quantifies the phenomenon of ASR models inflating scores by learning benchmark patterns rather than actual speech.
Applying iterative pseudo-labeling significantly improved the accuracy of Mandarin-English code-switching speech recognition.
Microsoft is testing MAI Realtime, a full-duplex real-time voice model, in MAI Playground.
VibeVoice-ASR-BitNet, an ultra-lightweight speech recognition model optimized for edge CPU environments, has been released.
This analyzes the role of benchmark datasets, a key factor determining the advancement and performance evaluation of voice-based AI technology, along with the current status in Korea and abroad.
This piece analyzes the advancement of speech recognition technology and the importance of benchmark datasets for measuring performance, comparing the status of domestic and international datasets.
Hume AI has launched Real World VoiceEQ, a new benchmark that evaluates emotional understanding and naturalness in voice AI.
Apple's new SpeechAnalyzer API overwhelmed on-device speech recognition performance, recording higher accuracy and faster speed than Whisper models.
A 0.9B-scale audio understanding model that performs speech transcription, speaker diarization, and timestamp generation all at once has been released.
The paper proposes four improvements to address the loss of positional information that occurs when AED models process long audio.