Tag
Krisp released an open benchmark and dataset showing that voice isolation reduces word error rates in speech-to-text models by 73%, with significant improvements across workplace and call-center recordings.
Wispr Flow introduces Canto, a speech model for real-time dictation that excels in noisy and challenging real-world conditions, achieving the lowest word error rate in evaluations against competitors.
Meta has released Muse Voice Transcribe, a real-time speech-to-text model with a 3.1% word error rate and adaptive streaming capabilities, making it suitable for voice agents.
This preregistered ablation study tests prompt-level context in a production speech transcription tool and finds no detectable change in side-level word error rate, contradicting earlier reports of gains from prompt conditioning.
This paper presents a controlled benchmark comparing six multilingual pre-trained ASR models on Nepali speech, finding Whisper-Large-v3-Turbo and IndicWav2Vec perform best, while CTC decoders offer up to 29x faster inference. It provides the first standardized efficiency-aware reference numbers for Nepali ASR.
This paper compares context biasing methods and speech LLMs for recognizing rare and new words in automatic speech recognition, reporting trade-offs in word error rate across read and non-read speech.
Room reverberation and low-frequency noise from the environment hurt speech-to-text accuracy far more than the choice of model size; front-end audio preprocessing like adaptive spectral subtraction can recover masked phonemes and reduce word error rate more effectively than upgrading the model backend.
Apple's new SpeechAnalyzer API significantly outperforms both its predecessor SFSpeechRecognizer and OpenAI's Whisper models in accuracy and speed for English on-device transcription, as benchmarked on an M2 Pro machine. The new API achieves a 2.12% word error rate on clean speech, compared to 3.74% for Whisper Small, and runs three times faster.
This paper revisits the classic relation between language model perplexity and ASR word error rate in the context of modern end-to-end ASR systems, finding that while external LMs still improve WER, the log-log linear relation still holds but is affected by internal language modeling in encoder-decoder models.
This paper presents a certification-inspired mechanism for automatic speech recognition that uses a dual-gate diagnostic pipeline (Two-Sided Atomic Audit and Rank-Based Tournament) to provide certified robustness and achieve up to a 55% relative reduction in word error rate across diverse architectures.
A speech company trained a model that cancels noise and identifies the primary speaker, achieving 50% lower word error rate on leading ASR models in noisy environments.
This paper critiques the use of single-reference ground truth in ASR evaluation, arguing it causes epistemic injustice for speakers with aphasia. It proposes a new metric, Epistemic Injustice Distance, and advocates for WER-Range to account for diverse transcription conventions.