A hotword is a bias, not an unconditional instruction to replace every similar sound. If ordinary words begin turning into a brand name, inspect the strength and scope of that bias. Alibaba’s ASR guidance warns that excessive weighting can cause unwanted substitutions. Increasing the weight further can make the complaint worse.
Preserve the words that should not change
Start with the original audio and a human-reviewed transcript. Mark both the intended brand mentions and nearby ordinary phrases. Without that baseline, a reviewer may celebrate a corrected brand name while missing new errors elsewhere. Check what is audible rather than assuming every similar syllable refers to the brand.
In a hypothetical workshop recording, an uncommon product name resembles a common phrase. A strong vocabulary entry might improve the product introduction while damaging an unrelated instruction later. The evaluation must include both. This is an authored test design; it is not a claim that a particular recognizer produced those errors.
Build positive and negative controls
Choose a short target passage containing a genuine mention and several control passages that should remain ordinary language. Include representative background conditions and speaking styles from the project. Do not create only crystal-clear studio controls for a noisy live recording, or use completely different voices when trying to compare one setting.
| Passage | Human-approved text | Desired behavior | Reject when |
|---|---|---|---|
| Genuine brand mention | Exact authorized spelling | Recognize the name | Name remains wrong |
| Similar ordinary phrase | Normal wording | Preserve ordinary words | Becomes brand name |
| Numbers and negation | Exact meaning | No collateral change | Qualification changes |
| Unrelated passage | Baseline text | Stable result | New false substitutions |
The table is an original acceptance plan, not a numerical accuracy score. Record actual findings and avoid claiming a universal safe weight.
Adjust only after application is verified
Confirm that the correct vocabulary is applied to a compatible recognition model before tuning it. Then consider a lower supported weight, removing irrelevant terms or narrowing the vocabulary to the project. Keep one meaningful change per comparison where practical. An overloaded word list can make the reason for a changed result difficult to identify.
If no setting preserves both the target name and ordinary speech, retain manual review or choose another suitable route. Do not trade a single brand correction for damaged instructions or false claims. Keep the final approved transcript independent from the raw recognition result so errors do not flow directly into published subtitles.
Approve the vocabulary for its real scope
Record which recording and project the vocabulary was tested against. A list suitable for one campaign is not automatically suitable for every future film. Review again after adding terms, changing audio conditions or switching models. Keep rejected settings out of shared defaults.
Imagild projects may use an externally checked transcript; this article does not claim an in-product hotword control. Its practical outcome is a vocabulary decision supported by both intended corrections and preserved ordinary language, followed by a full content check before delivery.