A vocabulary can be created successfully yet have no effect on a particular transcription. Before increasing weights, check that the task actually references the correct vocabulary and that its target model is compatible. Alibaba’s ASR accuracy guide documents this relationship. An available vocabulary in a management page is not proof that a recognition request used it.
Follow the vocabulary into the task
Record the vocabulary identifier, its intended model, the current recognition model and the task’s vocabulary reference. Inspect the submitted task rather than an earlier successful example. Similar vocabulary names are insufficient; a draft or another region’s resource can be mistaken for the approved one.
Consider a hypothetical exhibition film containing an unusual brand name. The team updates the word list but continues submitting an old vocabulary identifier. Raising the new list’s weight cannot affect that request. Another case uses the updated identifier with an incompatible model. Both need a configuration investigation before pronunciation or source recording is blamed.
Build a binding receipt
Keep the content of the word list separate from evidence that it was applied. Confirm the provider’s current supported model and regional requirements. Do not assume that a vocabulary created for one ASR family works for another, or that a client library forwards every optional field automatically.
| Check | Evidence needed | Failure classification |
|---|---|---|
| Vocabulary identity | Actual resource identifier | Wrong or old list |
| Model compatibility | Current documented support | Unsupported pairing |
| Task binding | Submitted parameter or supported UI selection | List never applied |
| Target word | Approved spelling and audible sample | Content or audio problem |
This is an original diagnostic receipt, not a transcription benchmark. If the product does not expose request evidence, mark binding unverified rather than claiming it succeeded.
Test a short, representative passage
Once binding is established, use an existing permitted sample or a deliberately planned test containing the target word and nearby ordinary words. Have a human establish what is actually spoken. If comparisons are available, keep audio and recognition settings constant so the vocabulary is the meaningful variable. No new paid request is required to complete the input audit.
Do not approve the whole vocabulary because one word improves. Check neighboring words and ordinary phrases for unwanted substitutions. If the target remains wrong, investigate clarity, pronunciation, language settings and model behavior before escalating weights. A documented feature does not guarantee recognition of every name.
Carry the verified resource into caption production
Record the approved vocabulary version and binding evidence with the task result. When changing model, region, integration or word list, recheck compatibility and application. Review all captions for meaning, numbers and names even after a successful test.
For an Imagild film, a verified transcript can be an externally prepared editorial asset; this article does not claim that Imagild manages Alibaba vocabularies. Its pass condition is a supported vocabulary visibly associated with the intended recognition task, followed by a separate review of the actual words returned.