Average Word Error Rate (WER) accuracy across 9 languages
Measured against standard reference transcripts using industry-recognised test sets. English, Japanese, German, French, Korean, Spanish, Portuguese, Mandarin Chinese, and Dutch.
What "98.7% Accuracy" Means
Transcription accuracy is measured using Word Error Rate (WER), the industry-standard metric for speech recognition quality. WER counts the minimum number of word substitutions, deletions, and insertions needed to transform the system's transcript into the reference transcript, divided by the total number of words in the reference.
Formula
WER = (Substitutions + Deletions + Insertions) ÷ Total Reference Words
Accuracy = 1 − WER = 98.7% means WER of 1.3%
A WER of 1.3% means that on average, for every 100 words spoken, OneMeet produces 1.3 incorrect words. In a 60-minute university lecture containing approximately 9,000 words, this corresponds to roughly 117 word-level errors — typically minor substitutions (e.g., "affect" vs "effect") rather than complete word losses.
Test Conditions
All accuracy measurements were conducted under the following conditions:
| Parameter | Specification |
|---|---|
| Audio input | Microphone capture at 16kHz, 16-bit PCM |
| Environment | Quiet room (≤30 dB ambient noise) and moderate lecture-hall noise (40–55 dB) |
| Speaker profile | Native speaker, non-native speaker, and mixed-accent speakers per language |
| Content type | Academic lecture excerpts, business meeting recordings, structured monologue |
| Reference transcripts | Human-annotated ground-truth transcripts reviewed by two independent annotators |
| Test set size | Minimum 2 hours of audio per language |
| Measurement tool | NIST SCLITE, the standard scoring toolkit for ASR evaluation |
Languages Tested and Per-Language Results
OneMeet was tested across all nine supported languages. Results are reported as accuracy (1 − WER), rounded to one decimal place.
| Language | Accuracy | Notes |
|---|---|---|
| English | 99.2% | Broadest training data; highest accuracy |
| Spanish | 99.0% | Both Latin American and Castilian Spanish |
| Portuguese | 98.9% | Brazilian and European Portuguese |
| French | 98.8% | Including academic and formal register |
| German | 98.7% | Including compound nouns and Fachsprache |
| Mandarin Chinese | 98.6% | Simplified character output; character error rate |
| Korean | 98.5% | Including agglutinative morphology |
| Dutch | 98.4% | Including code-switching with English |
| Japanese | 98.3% | Including kanji, hiragana, katakana output |
The 98.7% figure reported in OneMeet marketing materials is the unweighted average across all nine languages under standard test conditions.
Comparison Baseline
OneMeet's accuracy was benchmarked against two reference points:
- Industry average for multilingual ASR (2024–2025): Published benchmarks for general-purpose multilingual speech recognition systems report average WER of 3–8% across similar language sets (corresponding to 92–97% accuracy). OneMeet's 1.3% WER represents a meaningful improvement over this baseline.
- Human transcription accuracy: Professional human transcribers typically achieve 99.0–99.5% accuracy on clear audio. OneMeet's 98.7% average approaches human-level accuracy for clean audio conditions.
Known Limitations
Accuracy degrades in the following conditions. We report these transparently so users can set appropriate expectations:
- • Heavy background noise (>60 dB ambient): accuracy typically drops to 93–96%
- • Strong regional accents not well-represented in training data: accuracy can drop 2–4 percentage points
- • Code-switching mid-sentence (e.g., Dutch-English in a single sentence): partial accuracy reduction, typically 1–3%
- • Highly technical domain vocabulary not in the base model's vocabulary: proper nouns, very new terms, and niche technical jargon may be misrecognised
- • Low-quality microphones or lossy audio compression (e.g., .mp3 at <128kbps): accuracy reduction of up to 5 percentage points
- • Multiple simultaneous speakers: speaker diarisation accuracy is separate from transcription WER and varies by number of speakers
How We Continuously Improve
Accuracy metrics are reviewed quarterly. System updates are triggered when any language's accuracy drops below 98% on our internal test suite, or when a major model version becomes available. User-reported corrections contribute to our training pipeline through a privacy-preserving opt-in programme.
Using This Data
You are welcome to cite OneMeet's accuracy figures in academic work, journalism, or product comparisons. When citing, please reference this page and include the date you last accessed it. If you need raw WER numbers or test set details for research purposes, contact research@onemeet.ai.
Citation
OneMeet (2026). Transcription Accuracy Methodology. Retrieved from https://onemeet.ai/accuracy-methodology