Vocabulary mismatch drives most speech-recognition errors, says ASR provider

Speech recognition failures in production environments are primarily caused by vocabulary mismatch rather than poor audio quality or underperforming models, according to a statement released by an automatic speech recognition provider.

The company said that using constrained-vocabulary speech recognition can reduce the need for manual correction and help maintain consistent accuracy across different client implementations.

Runtime vocabulary customisation alone can deliver a 20–30% relative improvement over baseline accuracy, the company said. Fully optimised systems, with domain-specific model customisation, can reduce word error rates to between one and five percent.

The statement noted significant differences among vendors in support for custom vocabularies. Provider A allows up to 5,000 phrases per request, while Provider B limits inputs to 50KB files and a maximum of 100 vocabularies. Provider C recommends 500 phrases and reports warm-state latency of around 700 milliseconds. One provider was cited as having cold-start latency of up to 5,000 milliseconds, dropping to roughly 700 milliseconds when warmed. Most providers do not disclose quantitative data on latency overhead.

The company said runtime vocabulary injection avoids the need for tenant-specific models, provides natural isolation between clients—since vocabularies persist only during processing—and simplifies infrastructure scaling. Some teams, it said, accept modest accuracy trade-offs in return for operational simplicity.

Related reading

• Apple’s rumoured January event: Refresh, or reset?
• The new AI tech that aims to beat electricity grid constraints
• Microsoft unveils UniRG to improve accuracy in AI-generated radiology reports

The statement recommended customers begin with 500 to 1,000 essential terms per tenant, benchmark vocabulary sizes in the 2,000–5,000 token range using representative production audio, and audit vocabulary sets for phonetic confusability using transcriptions and Levenshtein distance metrics. It also advised testing on actual production hardware to ensure accuracy and performance.

To encourage evaluation, the company is offering $200 in credits for testing runtime keyword boosting features against existing providers.

The Recap

• Constrained vocabularies cut transcription errors versus general-purpose ASR.
• Runtime customization yields 20–30% relative accuracy improvement in production.
• Start with 500 to 1,000 core terms per customer.