New open-weight ASR model Whistle is released from Cactus Compute, weighing at only 16MB and performing on par with Whisper Base, enabling ultra fast on-device speech-to-text, huggingface.co/Cactus-Compu...

New open-weight ASR model Whistle is released from Cactus Compute, weighing at only 16MB and performing on par with Whisper Base, enabling ultra fast on-device speech-to-text, huggingface.co/Cactus-Compu...
Streaming ASR looks fluent but secretly gambles on every character. A new single-pass model decides when to commit vs. what to commit—no revisions, no global #SpeechRecognition #ASR #MachineLearning #AI
https://freegardner.com/synapse/machine-that-says-it-is-listening.html
Whisper adapted with LoRA on consumer GPUs handled three Arabic ASR subtasks at NADI 2026, reaching 57.1% WER on country-level ASR and 14.49% WER on Tunisian code-switched ASR.
#Whisper #LoRA #ASR #ArabicNLP
https://arxiv.org/abs/2610.09934
33 min and asr just dumped -13.59%, from 0.9974 straight down to 0.8619 👀 #ASR #crypto
🎙️ Can phase space dynamics boost ASR?
Most speech systems ignore signal phase. In our paper in SIVP, we use Recurrence Plots + 2D Adaptive Wavelets to capture non-linear dynamics.
📄 Read more: link.springer.com/article/10.1...
AI tools for transcription are getting popular, leading to productivity boosts from far less typing. I explore the free open-source options my latest from the blog reel:
I'm now using Whisper Clip on an hourly basis! I hope you find it useful.
Title: Non-lexical Sounds in ASR for Clinical Documentation
⚠️ NLCS poorly recognized in clinical ASR
🚫 Clinically relevant NLCS misrecognized with 94.7% (Google) and 98.7% (Amazon) error rates
bluesy
#ClinicalDocumentation #ASR #PatientSafety #NLP #HealthcareTech
Advantech Unveils ASR-D501 AI Companion Computer for Drones #Taiwan #Taipei #Autonomous_Drones #Advantech #ASR-D501
It highlighted how recognition errors can affect low-resource language speakers, older people, people with speech impairments, and accented English speakers differently.
These challenges are central to ALFIE’s work on inclusive and ethical AI.
半教師あり連合型ASRのための実践的レシピ:オンライン疑似ラベルとサーバー更新安定化
ASRにおける半教師あり連合学習の課題を克服し、実践的な安定化手法と疑似ラベル生成戦略を詳述。
#FederatedLearning #ASR #Semi-supervisedLearning #MachineLearning #Privacy
Really cool work from the NVIDIA team with a new speaker diarization model (Nemotron 3 Diarization). Since they gave me early access, I’ve been investigating the model and the Rust port for a while, and I released it the same day for parakeet-rs. 🙂
github.com/altunenes/pa...
#rustlang #asr #onnx
Amazon Elastic Compute Cloud (Amazon EC2) and NVIDIA collaborate with Heidi to reduce ASR inference costs by 75% using NVIDIA CUDA MPS. 🚀💸 #AWS #NVIDIA #HeidiHealth #ASR #InferenceCosts
Investigating non-lexical sounds in clinical ASR.
🔍 3284/135,647 words were NLCS
🗣️ 76 NLCS impacted clinical data
🧠 Google WER: 11.8%, Amazon: 12.8%
🔊 NLCS WER: Google 40.8%, Amazon 57.2%
#ASR #EHR #NLP #PatientSafety
https://tnyp.me/Npmiz0F4/s
Research explores NLCS impact on ASR in clinical docs.
🧩 135,647 words; 3284 NLCS (2.4%)
🧩 76 NLCS (2.3%) held clinical info
🧩 WER for Google ASR: 11.8%; Amazon ASR: 12.8%
#ClinicalDocumentation #ASR #NLCS #HealthcareInnovation
Voice Arena and Hugging Face team up for an open ASR evaluation in Hindi and Indian English. New benchmarks focus on regional variation and orthographic challenges. #ASR #VoiceRecognition #Hindi #IndianEnglish 🗣🇮🇳
Microsoft Defender ASR blocks suspicious Windows behavior, from WMI process creation to Office abuse, while logging config changes. Wazuh ingests ASR telemetry to detect blocked actions and ASR tampering on endpoints. #MicrosoftDefender #Wazuh #ASR