Word error rate
Also written: WER, Live-speech error rate, Real-time WER
Word error rate measures how many words a speech-to-text system gets wrong. Lower scores mean better accuracy.
Word error rate, or WER, is a way to measure how good a speech-to-text system is. Think of it like grading a student's spelling test. If the system transcribes "hello world" as "helloworl", that's one error.
WER is calculated by counting the number of mistakes a system makes (like misspellings, missing words, or extra words) and dividing it by the total number of words it was supposed to transcribe. A lower WER means the system is more accurate. For example, if a system transcribes a sentence with 10 words and makes 1 mistake, its WER is 10% (1 error / 10 words).
Many AI companies are working to improve their speech-to-text models, aiming for lower WER scores. This makes voice assistants, transcription services, and other audio-based AI tools more reliable and useful.
Word error rate (WER) is a standard metric for evaluating automatic speech recognition (ASR) systems. It quantifies the difference between a recognized word sequence and a reference word sequence. The calculation involves counting the minimum number of substitutions, deletions, and insertions required to transform the recognized text into the reference text, and then dividing this count by the total number of words in the reference text. A lower WER indicates higher accuracy. For instance, a WER of 0.05 means that, on average, 5% of the words in the transcription were incorrect.
Recent advancements in large language models and specialized audio models, like those from OpenAI and Google Gemini, are pushing the boundaries of speech recognition accuracy. Companies are reporting lower word error rates, making transcription more reliable for real-world applications. This focus on improving WER is crucial for the development of more sophisticated voice-enabled AI products and services.
- GPT-4o Transcribe ModelOpenAI · Model Documentation ·
- GPT-4o Mini Transcribe ModelOpenAI · Model Documentation ·
- Updates for developers building with voiceOpenAI · OpenAI Developers Blog ·
- Google announces Gemini 3.5 Transcribe for AI-powered speech-to-textGoogle · Ars Technica AI ·
- Intelligent transcription with Gemini 3.5 TranscribeGoogle · The Keyword: Gemini ·
- How developers build AI for good with Gemma 4Google · The Keyword: Developers ·
- Introducing next-generation audio models in the APIOpenAI · OpenAI News ·
- Google Gemini
- Google Gemini's transcription models aim for precise, intelligent real-time speech-to-text, implying a focus on minimizing word errors.
- OpenAI ChatGPT
- OpenAI's GPT-4o and Whisper models are being updated to deliver more accurate audio transcription, directly addressing word error rate improvements.
- Character Error Rate (CER)
- Measures errors at the character level, not the word
- BLEU Score
- Primarily used for evaluating machine translation quality by compari
- ROUGE Score
- Used for evaluating summarization and machine translation, focusing
Word error rate was developed in the context of early speech recognition research to provide a quantitative measure of performance. It has become a de facto standard across the field. Source ↗
Use a voice assistant like Google Assistant or Siri. Speak a short, clear sentence and then ask it to repeat what it heard. Compare its output to what you actually said. Many AI chat interfaces also have a voice input option where you can test this.
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- PromptRole-play a student struggling with math for AI tutor simulationSimulate a student's math struggles and uncertainty for AI tutor evaluation. · 5 min
- TemplateReusable workflow templates for healthcare teamsTeams can create and share templates for common workflows like documentation or prior authorizations. · 15 min
- TemplateTemplate for building an AI tool tailored to your workA structured approach to defining and building a custom AI tool for specific business needs. · 15 min
- TemplateEvals CSV prompt list for agent skillsA CSV file template for listing prompts used in evaluating agent skills. · 20 min
- HeuristicDouble coding agent task completion with docs linksProviding coding agents with a documentation link significantly increases their task completion rate. · 2 min
Questions people ask
- Is a WER of 0% possible?
- A WER of 0% means perfect transcription with no errors. This is extremely difficult to achieve in practice due to variations in speech, background noise, and accents.
- How does WER relate to transcription quality?
- A lower WER score indicates higher transcription quality. For example, a system with a 10% WER is generally considered better than one with a 20% WER.
First seen on Subvolts Mar 20, 2025. Updated Oct 5, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.