A custom vocabulary can help a recogniser with unfamiliar product names, but it is not an exact-output guarantee. Keep a reviewed list and test whether it improves the actual user's terms without creating false substitutions elsewhere.
Separate vocabulary support from later cleanup. A correct raw name can still be changed by rewriting, while an incorrect raw name may need explicit evidence rather than a confident guess.
Build a focused vocabulary record
Include exact spelling, expected spoken form, context, and confusable alternatives. Keep the list relevant to current work rather than adding every word the user might ever say.
Review the chosen provider's current official feature documentation. For example, Deepgram's keyterm reference describes a specific supported mechanism and its handling considerations. Other recognisers and models may differ, so do not assume identical syntax or guarantees.
Evaluate a fictional product-name set
Use harmless names such as CedarSync and CedarDesk in sentences that also contain ordinary similar-sounding words. Record the same authorised sample before and after the supported vocabulary configuration.
Compare raw transcript, cleaned output, and final insertion. Test unknown name, plural or possessive form, correction, another language around the name, and a sentence where the ordinary word is intended instead of the product.
Count critical-name errors and false substitutions separately from overall readability. No improvement percentage or successful provider result is claimed in this proposed exercise.
Keep vocabulary data under the same privacy review
Names and project terms can be sensitive even when they are not full transcripts. Understand where the vocabulary is sent, stored, logged, and retained under the actual provider and application design.
Offer clear review and removal of outdated terms where the product supports it. A stale company name or confidential client code should not remain in a broad shared vocabulary indefinitely without a reason.
The tradeoff is contextual help versus overbias and maintenance. A focused list can support current terminology while an excessive list may add ambiguity and review work. Evaluate realistic sentences rather than isolated name repetition alone.
Preserve a manual exact-text path for critical strings. A spelling correction should remain deliberate, especially before executing code, sending a client message, or identifying an account.
The Mac dictation guide and StackSpeech's portfolio entry provide related context. This vocabulary plan does not claim StackSpeech currently exposes a custom-vocabulary feature or guarantees exact names.
Questions about custom dictation vocabulary
Does a vocabulary force the exact spelling?
Do not assume so. Verify the recogniser's current contract and evaluate the actual output and cleanup stages.
Should all project names be sent to a provider?
Only under an appropriate authorised data-flow policy. Minimise the list to the task and review sensitive terms deliberately.
What should be tested first?
Compare a shared sample with and without a small term list, including confusable ordinary words. Bring false substitutions to a voice-input evaluation review.