Scholarlink

With the rise of Automatic Speech Recognition (ASR) systems, it becomes imperative to analyze the environmental impact of such systems. This work presents a comparative study of energy consumption, carbon emissions, and performance in OpenAI's Whisper (base and large) versus Google Speech-to-Text V2 (Chirp/USM). In particular, the optimized implementations are considered: Whisper models with faster-whisper inference engine and Google STT V2 with dynamic batching. Operational energy and emissions measurements for an experiment involving 20,000 short Urdu audio clips (~22 h) were performed using CodeCarbon, PyJoule, PowerAPI, and Google Cloud Carbon Footprint. The results show that Google STT V2 with dynamic batching is the winner concerning speed and energy efficiency. Whisper large (faster-whisper) gives the highest accuracy but at the cost of energy and carbon emissions. The discussion revolves around trade-offs between accuracy, speed, sustainability, privacy, and cost, giving recommendations for sustainable ASR deployment. This study, therefore, underlines the importance of software optimization, life cycle assessment, and low-resource language support.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Sustainable Speech Recognition: Energy, Carbon, and Performance Comparison of Whisper (Base and Large) and Google Speech-to-Text V2 (Chirp/USM)

  • 章节

  • 发表日期：

    2026年

查看全文
  • Jalal El Bahri,
  • Mohamed Kouissi,
  • Mohammed Achkari Begdouri

摘要

With the rise of Automatic Speech Recognition (ASR) systems, it becomes imperative to analyze the environmental impact of such systems. This work presents a comparative study of energy consumption, carbon emissions, and performance in OpenAI's Whisper (base and large) versus Google Speech-to-Text V2 (Chirp/USM). In particular, the optimized implementations are considered: Whisper models with faster-whisper inference engine and Google STT V2 with dynamic batching. Operational energy and emissions measurements for an experiment involving 20,000 short Urdu audio clips (~22 h) were performed using CodeCarbon, PyJoule, PowerAPI, and Google Cloud Carbon Footprint. The results show that Google STT V2 with dynamic batching is the winner concerning speed and energy efficiency. Whisper large (faster-whisper) gives the highest accuracy but at the cost of energy and carbon emissions. The discussion revolves around trade-offs between accuracy, speed, sustainability, privacy, and cost, giving recommendations for sustainable ASR deployment. This study, therefore, underlines the importance of software optimization, life cycle assessment, and low-resource language support.