soundsLike "a"?

[VaunMcCarthy]

Hi guys

When using the API for building dictionary terms, what's the correct soundsLike for the letter A where it's pronounced as in USA? would "a" on it's own be more like as in CubA? @Leor is there a reference somewhere for consonants, vowels etc that might help here?

[Leor]

Hi Vaun! Thanks for this... checking and hope to update shortly.

Leor

[Leor]

Hello Vaun,

The feedback from the research team as to use the letter itself for the soundsLike, e.g. "USA" would be "u s a". The best way to utilize the soundsLike would be to see how the error looks in the transcription and to input that as the soundsLike example.

Hope this helps and happy to dive in more.

Leor

[VaunMcCarthy]

Thanks Leor I had thought of that but the examples given for IRS:

• 1. Eye are ess
• 2. Eye Are es
• 3. I R S
• 4. Ire rest
• 5. I RS
• 6. IR S
• 7. Aye heiress
• 8. I heiress

Seem to imply there's a need sometimes for additional soundsLike entries (understandable). I had considered something like the Canadian "eh" but people in some dialects would actually pronounce that like the e in get

[Leor]

Hi Vaun! Yes, the gold examples are the ones that are actually in the transcripts but the additional soundsLike examples should only positively affect the accuracy as they're updating the language model.

[VaunMcCarthy]

Hi Leor

There's a bit of confusing here hopefully you can clear up. What impact does the topic spotting configuration side of things have on transcription/summary accuracy? Everything seems to point to just needing the dictionary, but are there scenarios where topic phrase tuning should also be done?

[Leor]

Hi Vaun,

Topic Spotting has a side benefit of improving transcription for the phrases used to find the topic.

If you add a phrase, it biases the transcription engine to find that phrase vs other random text in the voice transcription. The drawback of using Topic Spotting for improving accuracy is that you will highlight many more topics and it dilutes the impact of spotting those topics you want to use to understand better what's happening in an interaction.

This is why we now suggest using Dictionary Feedback. It not only improves the accuracy for specific terms, but has the added ability to correct mis-transcriptions (Sounds Like), and will not highlight text in the transcript like Topic Spotting. The API version also has a "boost" value, which can further tune the transcription engine.

Short answer - yes, just use Dictionary Management for improving transcription accuracy.

Very grateful for the question.
Leor