What Is Word Error Rate (WER)?
Word error rate (WER) scores speech recognition by comparing its output to a correct transcript and counting substitutions, deletions and insertions.
إجابة سريعة
Word error rate (WER) is the standard way to score speech-to-text accuracy. The recognizer's output is aligned with a correct reference transcript, and the substituted, deleted and inserted words are counted as a share of the words in the reference. Lower is better, and the figure only means something for the audio, language and vocabulary it was measured on.
When any transcription product is described as a certain percentage "accurate", the useful question is: measured how, on what audio. Word error rate is the usual answer to the first half. It turns the comparison between what was said and what the software wrote into one number, and NIST publishes a scoring toolkit that implements it.
The Definition
The reference implementation most often cited is sclite, part of the NIST Scoring Toolkit (SCTK). Its documentation describes it as a tool for scoring and evaluating the output of speech recognition systems. It compares the hypothesized text (HYP) produced by the recognizer with the correct, or reference, text (REF). Before counting, the two are aligned. sclite can align with dynamic programming, which its documentation calls the first and most widely accepted method, or with GNU diff. By default, the dynamic programming alignment uses word-to-word distances of 0 for a correct word, 3 for an insertion, 3 for a deletion and 4 for a substitution, and performs a global minimization of that distance, so the alignment chooses the cheapest explanation of the differences.
After alignment, the errors are tallied. sclite reports the percent of correct words, substituted words, inserted words and deleted words, each as a share of the number of reference words. In the sample summary reports in the SCTK documentation, the total error column (Err) is the substitution, deletion and insertion figures added together, and that total is what is usually called the word error rate. Because inserted words are counted against a reference that did not contain them, the total can in principle exceed 100 percent in a very bad case.
كيف يعمل
- Substitution. A different word was written: "fifteen" for "fifty".
- Deletion. A word that was said is missing from the output.
- Insertion. A word appears that nobody said.
A Worked Example
This is an invented example. The reference transcript has 10 words. The recognizer substitutes one word, drops one word and adds one extra word. That is 1 substitution, 1 deletion and 1 insertion, 3 errors in all against 10 reference words, so the WER is 30 percent. The same three mistakes in a 100-word transcript would score 3 percent, which is why a WER is only comparable when the amount of reference text is large enough to be representative.
Why It Matters for Outbound and Inbound Teams
A contact center uses transcripts for summaries, keyword flags, quality review and search, so the accuracy of the words under them sets a ceiling on everything built on top. A missed negation or a wrong number in a payment promise matters more than a dropped filler word, but WER treats all errors equally by default. The sclite documentation says exactly that: the default scoring treats word errors with equal weight, with a separate weighted-word scoring as an option.
DialerBee's نسخ المكالمات separates the agent and customer audio channels and transcribes each speaker independently, detects the language automatically and supports 11 languages with dialect awareness. This page gives no DialerBee WER figure, and none should be assumed. A claim of accuracy should come with a test on your own recordings. The same applies to تحليل الكلام built on transcripts and to any conversational AI that listens before it answers. If your calls are in more than one language, see the comparison of multilingual and English-only AI.
How to Test a Vendor Fairly
- Collect a sample of your own recordings: real phone audio, your accents, your products and names.
- Have a person produce the reference transcripts.
- Run every candidate on the same audio and score all of them the same way.
- Report WER per language and per call type, not as a single blended number.
- Look at the important words separately: numbers, names, amounts and dates.
Limits of the Metric
- It depends on the audio. Noisy or compressed calls score worse than clean studio audio.
- It depends on the language. The sclite documentation notes that for languages such as Mandarin, character error rate is considered the primary performance measure.
- It is not meaning. Two outputs with the same WER can differ greatly in how useful they are.
الأسئلة الشائعة
What is a good word error rate?
Lower is better, but a good figure depends on the audio, the language and the vocabulary. Compare systems only on the same test recordings, and judge them on your own calls.
How is word error rate calculated?
Align the recognizer output with a reference transcript, count the substituted, deleted and inserted words, and express each as a share of the reference word count. Their sum is the WER. For example, 3 errors in a 10-word reference is 30 percent.
Can word error rate be above 100 percent?
Yes, in principle. Inserted words add errors without adding reference words, so a poor output with many insertions can exceed the length of the reference.
Is WER the right measure for every language?
Not always. The sclite documentation notes that for languages such as Mandarin, character error rate is considered the primary measure.
Does a low WER mean the transcript is useful?
Not necessarily. WER counts words equally, while for a business use a wrong number or name matters far more than a small filler word. Check the important words separately.
Related terms and guides: نسخ المكالمات, تحليل الكلام, conversational AI, multilingual vs English-only AI.
مصادر
- NIST Speech Recognition Scoring Toolkit (SCTK) (source, checked November 7, 2026)
- sclite documentation, SCTK (alignment distances, word and weighted-word scoring, the Mandarin note) (source, checked November 7, 2026)
- sclite output reports, SCTK (sample summary with Sub, Del, Ins and Err columns) (source, checked November 7, 2026)
- NIST Multimodal Information Group, tools (lists the speech recognition scoring toolkit) (source, checked November 7, 2026)
قراءات ذات صلة: كيفية قياس دقة التفريغ النصي للمكالمات (WER) · المزج بين العربية والإنجليزية في الذكاء الاصطناعي الصوتي والتفريغ النصي
قراءات ذات صلة: ما هو التعرّف التلقائي على الكلام (ASR)؟
مصطلحات ذات صلة
ما هو مركز اتصال خدمات التعهيد الخارجي (BPO)؟
مركز اتصال BPO يدير عمليات الاتصال نيابةً عن شركات عميلة. ما يتطلبه النموذج: عزل المستأجرين، قواعد لكل عميل، علامة تجارية، وتقارير.
ما هو سجل تفاصيل المكالمات (CDR)؟
A call detail record logs who called whom, when, how long and how the call ended. See what a CDR holds, how teams use it, and how it differs from a recording.
ما هو تصنيف المكالمة؟
نتيجة المكالمة هي النتيجة المشفرة التي يسجلها الموظف في نهاية الاتصال. تعرّف على كيفية تصميم الرموز، وأسباب تغيرها، وما الذي يعتمد عليها.
What Is a Data Processing Agreement (DPA)?
A data processing agreement (DPA) is the contract GDPR Article 28 requires between a controller and a processor. Learn what it must cover for call recordings.
ما هو رقم DID (الاتصال المباشر الداخلي)؟
رقم DID هو رقم هاتف يُوجّه المكالمات مباشرةً إلى وجهة داخل نظامك. تعرّف على كيفية عمل مخزون أرقام DID ، ووجهات التوجيه، وإدارة الأرقام.
ما هي قائمة عدم الاتصال (DNC)؟
قائمة "عدم الاتصال" (DNC) هي سجل لأرقام الهواتف التي اختارت عدم تلقي مكالمات التسويق عبر الهاتف. تعرّف على أنواع قوائم "عدم الاتصال"، وكيفية تنظيفها، وحظرها.
هل أنت مستعد لرؤية تطبيق DialerBee أثناء العمل؟
اطلب عرضًا توضيحيًا مباشرًا مدته 15 دقيقة، أو ابدأ تجربة مجانية لمدة 14 يومًا — دون شرائح عرض، ودون أي التزام. يبدأ الاتصال بمجرد ربط خط SIP أو نظام PBX الخاص بك واعتماده.
تجربة مجانية لمدة 14 يومًا · رصيد ذكاء اصطناعي مجاني · بدون بطاقة · 11 لغة · BYOC · ضوابط داعمة للامتثال