WER

Word Error Rate (WER) is the most universal performance metric in the speech recognition field, calculated as (number of substitutions + number of deletions + number of insertions) / total number of words in the reference text.

Gemini 3.5 Transcribe is an extension of the Gemini series in the direction of speech processing, with two core capabilities of "precision" and "intelligence" as its main features.

The mainstream metric for evaluating speech naturalness is Mean Opinion Score (MOS), where human listeners rate samples on a scale of 1 to 5.

The formula for Word Error Rate (WER) is (substitutions + deletions + insertions) / total number of words in the reference text, and a lower WER indicates higher transcription quality.

The word error rate formula is WER = (S + D + I) / N, where S is the number of substitution errors, D is the number of deletion errors, I is the number of insertion errors, and N is the total number of words in the reference text.

Voice drift is a known technical challenge in autoregressive speech generation models, where accumulated errors during step-by-step generation cause the timbre to deviate from the target speaker.

Children's shorter vocal tracts result in higher and more variable formant frequencies, which is one of the reasons for the difficulty in children's ASR recognition.

Current state-of-the-art systems have achieved a WER below 5% on the standard English test set LibriSpeech.

Early mainstream speech recognition solutions relied on Hidden Markov Models (HMM), which had limited capabilities for noise, accent, and multilingual mixed recognition.

Gemini 3.5 Transcribe can automatically remove false starts and filler words, and identifies self-corrections to retain only the final result.

The ASR Round-Trip approach converts synthesized audio back into text and uses edit distance or word error rate (WER) to measure the deviation from the original text, which can directly capture pronunciation content errors.

In detention facility scenarios, ASR faces the core challenges of far-field recording and channel degradation, including low sampling rates, limited frequency response, accent differences and multilingual mixing, which significantly drive up the word error rate (WER).

In voice agent scenarios, neither WER nor MOS is sufficient: WER only covers comprehension and does not reflect intent understanding; MOS only measures the naturalness of single-sentence synthesis and does not measure multi-turn dialogue context consistency or emotional expression.

WER (word error rate) measures the difference between ASR transcription and reference text, and MOS (mean opinion score) measures the naturalness of synthesized speech; these were the two core metrics for speech evaluation in the past.

WER is calculated as (number of substituted words + number of deleted words + number of inserted words) / total number of reference words, and is extremely sensitive to text formatting (case, punctuation, abbreviations, spelling variants), so normalization is usually required before evaluation.

Word Error Rate (WER) is calculated by dividing the sum of substitution, deletion, and insertion errors between the recognition output and the reference text by the total number of words in the reference text.

Falcon-ASR achieved an average Word Error Rate (WER) of 20.92% across 6 Arabic test sets.

The WER of top-tier ASR systems for English can be as low as below 5%.

Academia typically measures OCR quality using Character Error Rate (CER) and Word Error Rate (WER), and these metrics have already approached human-level performance on standard printed-text datasets.

For encoder-decoder models, direct bilingual fine-tuning achieves the strongest performance on the raw WER metric.