VoiceInk Docs
Dictionary

VoiceInk Vocabulary Accuracy Study

How often the local vocabulary matcher recovered preferred name spellings, and how often it changed text that should stay the same.

In VoiceInk's October 2026 vocabulary study, the local matcher recovered the preferred spelling in 1,805 of 2,000 full-name spelling-variant cases (90.25%) with 128 vocabulary entries. In separate no-change controls at that dictionary size, it made 6 unwanted edits in 3,000 cases (0.2%). This was an evaluation of the text correction step after transcription. It used prepared text inputs and measured exact whole-sentence matches.

Name correction results

All rows below used dictionaries with 128 entries.

Input typeExact sentence recovery
Full-name spelling variant1,805 of 2,000 (90.25%)
One part of a name varied578 of 1,000 (57.8%)
Full-name homophone substitution945 of 1,000 (94.5%)
Single-part homophone substitution125 of 500 (25.0%)
Name variant with a different number of words91 of 500 (18.2%)

Results varied sharply by input type. The matcher was more effective at recovering full names than a single component or a variant with a different number of words. If there is a recurring error that should always map to a known spelling, use a Word Replacement.

How dictionary size affected mistakes

Vocabulary entriesExact recovery on correction casesUnwanted edits on no-change controls
15,461 of 7,2160 of 3,000
325,563 of 7,3662 of 3,000
1285,557 of 7,3666 of 3,000
5125,526 of 7,36624 of 3,000

The one-entry condition excludes 150 cases that require two target names. Across the matching 32-, 128-, and 512-entry conditions, the larger dictionaries produced more unwanted edits in this sample. Start with names and terms you actually dictate.

Method and limits

VoiceInk ran the evaluation on October 27, 2026 using the unmodified production vocabulary matcher at the time. The evaluation contained 10,371 text cases and generated 41,319 predictions across dictionary sizes and diagnostic cases. No thresholds were tuned for this study.

  • We sampled observed spelling variants from JRC Names, excluding 2,718 entities used in an earlier evaluation. An observed spelling was supplied as the preferred dictionary entry.
  • We used CMUdict to make controlled homophone substitutions independently of VoiceInk's matcher. The variants were embedded in generated sentences.
  • No-change controls included already-correct names, names absent from the dictionary, structured literals, and unchanged web sentences from Universal Dependencies English EWT.

The study measured exact text recovery. It did not measure microphone capture, speech recognition accuracy, word error rate, end-to-end latency, or competing apps. The source names can include legitimate alternative spellings and some automatically grouped variants may be wrong. CMUdict mainly represents US English, and the evaluated matcher did not recover the cases in the unsupported accent and punctuation diagnostic sets. VoiceInk developed the feature and ran the evaluation.

See Vocabulary for how to add names and technical terms to your own dictionary.