Earlybird predicts spoken words before they end in 29M parameters
On LibriSpeech test-clean the median word is correct 90 ms early, and 69% of words are right before they end.
Stories per day · 30 days
3 RSS
Relatedopen-weights4huggingface2nemotron2nvidia2streaming-asr2turkish2voice-agents2apache-2.01apple1arabic1asr1audio-language-model1bemba1bias-evaluation1decision-models1dialogue1earlybird1elderly-speech1
30 stories, newest first · one card per story
Stories tagged #speech-recognition
Telefónica submits ASR system for Speech-COSER Challenge 2026
An ensemble of seven LoRA systems reached 22.17% normalized word error rate on elderly rural Spanish speech.
Confucius4-R2T2 released as stable-prefix streaming ASR model
R2T2 maintains stable prefixes in real-time transcription without modifying earlier outputs while open-sourced by Netease Youdao
GigaSpeech 3 for Arabic speech recognition
39,000-hour GigaSpeech 3 corpus covers five Arabic dialects and reduces WER by 16.6% relative to prior systems.
ufakzeka-kulak Turkish speech recognizer released
Open-source 31.6M-parameter Turkish speech-to-text model outperforms Whisper base on FLEURS and Common Voice tests.
TurnTurk-GPT predicts Turkish turn-taking from unpunctuated text
The fine-tuned Turkish GPT-2 beats a multilingual detector, but ties a zero-shot model on unseen speakers.
VoxLock extracts a target speaker for voice agents without enrollment
On Krisp's 265-recording set it reaches 4.95% WER versus 23.21% unprocessed, with 24 ms latency.
NVIDIA releases open Nemotron 3.5 Transcribe, a 30B MoE speech model
The open model transcribes 40 languages while activating about 4 billion of its 30 billion parameters.
Hugging Face introduces AnySpeechRecognizer and AnyDecisionModel
The announcement covers open models for speech recognition and decisions on Apple platforms, with no further details available.
Nvidia releases open Nemotron 3.5 Transcribe speech model
Nvidia says the 30-billion-parameter model keeps about 4 billion active and tops open-weight English ASR models.
GRPO treats ASR pseudo-labels as rewards rather than targets
In banking audio, GRPO cut WER 29.3% to 56.2% relative to SFT and copied 27% of teacher errors.
IBM announces Granite Speech 5.0 for speech tasks
The Hugging Face post title says the model transcribes, translates, and attributes speech more accurately and faster.
Qwen-Audio-3.1-ASR adds diarization and broader audio understanding
The family keeps 30 languages and 16 Chinese dialects, with file, streaming, and interactive variants.
AfriBED evaluates gender and speaker-type bias in Bemba and Nyanja ASR
The set holds about 9 hours of Bemba and 18 hours of Nyanja from 80 and 127 speakers.
TII Releases Falcon-ASR: 1.6B Arabic Speech Model Hits 20.92% WER
TII released Falcon-ASR, a 1.6B Arabic speech model reporting 20.92% average WER; weights are not public.
Oído: Open-Vocabulary Speech Recognition on a $5 Microcontroller without a Neural Accelerator
Oído runs open-vocabulary English speech recognition on a $5 ESP32-S3 without a neural accelerator.
Introducing Falcon ASR
TII introduced Falcon ASR, a speech-recognition release in the Falcon model family.
FermionResearch/Phonon-2 — new model trending #29 on Hugging Face
Fermion Research released Phonon-2, a 164 MB English speech model averaging 5.21% word error.
3
Pruned CTC for Memory-Efficient Large-Vocabulary ASR Training
Pruned CTC restricts alignment computation to target tokens and blank, cutting ASR training memory 5.1x for 180K-vocabulary LLMs with minimal accuracy loss.
A Training Criterion with Token-Level Tolerance to Transcription Ambiguity for Automatic Speech Recognition
Token-level wildcard paths in CTC cut ASR word error by tolerating localized transcript ambiguity.
Edge0/Audio8-ASR-Infinite — new model trending #30 on Hugging Face
Edge0 released Audio8-ASR-Infinite, a bilingual streaming speech model with constant-memory unlimited transcription.
8
Design of the IBM Granite 5.0 TurboCTC ASR Model
IBM details Granite 5.0 TurboCTC, a 470M-parameter Conformer-based English ASR model on the Open ASR speed-accuracy Pareto frontier, released on Hugging Face.
Iceland-based Treble raises $18 million for its voice simulation platform
Iceland-based startup Treble raised an $18 million Series A extension to expand its acoustic simulation and synthetic data platform for voice AI companies.
AI for everyone in every language
Google says its AI now spans 300+ languages reaching 7 billion people, unveiling Gemini 3.5 Transcribe, Live Translate, and TranslateGemma models.