At the travel assistant of Nordlicht Reisen GmbH, voice messages and calls arrive (R-04), and answers are also to be spoken (R-05). Customers speak German, English and Danish (R-01). This lesson chooses the fitting path of Azure Speech in Foundry Tools for each load shape and shows the .NET side: file transcription, real-time recognition with the Speech SDK, speech output and the question of custom speech models.
Prior knowledge. The row "speech to text and back" of the decision between model and Foundry Tool is in lesson 1.1; keyless sign-in in lesson 1.4. There is no AI track lesson on speech. The telephone variant with Voice Live is the subject of lesson 4.4.
New in this lesson. Voice messages are files of at most five minutes (assumption). Call recordings arrive in large amounts at night. The editorial team maintains a list of place names (Skagen, Bornholm, Funen, made up as a selection) that the recognition must hit.
Paper, account, costs. Paper: decision and exercises; for a run, the examples need a Speech resource and recordings of your own, made up (R-12). Cost drivers: audio time for recognition, speech output by amount of text, for Custom Speech additionally training and hosting of the endpoint. The units per voice type are on the pricing page and were not read: open.
The requirement
All numbers are assumptions of the task.
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 4.3-01 | A voice message is shown as text in the portal. | F | file up to 5 minutes, result in under 30 seconds | Must | R-04 |
| 4.3-02 | Call recordings are transcribed at night with speaker assignment. | F | 500 files per night, no real-time pressure | Should | R-04 |
| 4.3-03 | Answers can be retrieved as speech in three languages. | F | 3 voices, 1 per language | Must | R-05 |
| 4.3-04 | Place names from the editorial team's list are recognized correctly. | Q | at most 10% errors on the list in the test set | Should | R-10 |
| 4.3-05 | No key in code and configuration. | R | 0 | Must | R-07 |
| 4.3-06 | Recordings and texts are made up and deleted after a deadline. | R | deadline in the configuration | Must | R-12 |