AI Briefing#asr
The latest AI and developer news about #asr, with the original source and a short summary.
The latest AI and developer news about #asr, with the original source and a short summary.
Superwhisper has released S1-mini, a 0.6B parameter text normalization model that runs on CPU.
Applying iterative pseudo-labeling significantly improved the accuracy of Mandarin-English code-switching speech recognition.
A provenance error was observed in ChatGPT's audio preprocessing pipeline, where generated text is treated as direct user input.
NVIDIA supports on-device execution by quantizing ASR, TTS, and speech codecs to GGUF.
Medical STT reduces the workload of medical staff by converting spoken clinical conversations into accurate medical records through recognition of clinical terminology.
VibeVoice-ASR-BitNet, an ultra-lightweight speech recognition model optimized for edge CPU environments, has been released.
audio.cpp 0.4 adds Higgs Audio v3 TTS and Fish Audio S2 Pro among others, now supporting 35 model families with Q8 GGUF improving speed and VRAM.
Word Error Rate (WER) is a key metric for measuring the accuracy of speech recognition (ASR) models, evaluating performance through insertion, deletion, and substitution errors.
A high-performance, multi-model speech recognition (ASR) library built on ggml has been released.
Cohere has released an open-source Arabic speech recognition (ASR) model with accuracy that surpasses Whisper.
This work proposes a method that improves the efficiency and accuracy of ASR error correction by leveraging a seq2seq model lighter than an LLM.
audio.cpp, a ggml-based C++ audio inference framework, has been released, offering up to 5x faster speed compared to Python.
Hugging Face and Treble Technologies have launched the FFASR leaderboard, which reflects real-life acoustic environments.
A new benchmark and dataset have been released to evaluate speech recognition performance on code-switching speech that mixes multiple languages.
NVIDIA has unveiled Nemotron 3.5 ASR, a roughly 600-million-parameter model that supports 40 languages in real time.
A new approach was unveiled that achieves low-latency multilingual speech recognition by routing single-language models.
It is a technology that achieves high-performance multilingual speech recognition in local environments by routing small monolingual models.
It's a CLI tool that converts clean speech data into noisy data resembling real call environments to test ASR performance.
vibevoice.cpp, a ggml-based inference engine that can run Microsoft VibeVoice without Python, has been released.
Microsoft released VibeVoice, unveiling a family of ASR, TTS, and real-time voice models.
Microsoft's VibeVoice transcribed 1 hour of audio in 8 minutes 45 seconds.
Xiaomi MiMo released an 8B ASR model that is robust to multiple languages, dialects, and noisy environments.
On-device streaming ASR achieved **8.20% WER** and **0.56s** latency on CPU.
An arXiv paper presents a lightweight English model for low-latency on-device streaming ASR.