Analysis portal overview
The Analysis Portal is a web-based tool for measuring and tuning the performance of speech applications running on Capacity Private Cloud. It is integrated with the Deployment Portal and is intended for tenant-level access only, operating on speech data within a single deployment.
The portal supports a complete tuning workflow — from gathering production transactions, through manual transcription, to detailed accuracy reporting — allowing deployment administrators to continuously improve recognition quality without leaving the browser.
Workflow
- Create Analysis Sets — Build a working set from existing transactions by filtering on date range and interaction type. A pre-create preview reports the number of sessions and interactions that match your filter before the set is generated, so you can size the job appropriately. Supported interaction types include ASR, CPA, AMD, Transcription and TTS.
- Transcribe — Establish ground truth by annotating each interaction by hand. The transcriber includes an embedded audio player with autoplay, rewind/advance controls, and a decoded view that overlays the Engine's recognition output on the audio timeline. Keyboard shortcuts (Enter to save and advance, Escape to mark Out-of-Coverage, arrow keys to navigate) speed up bulk transcription. A noise-marker toolbar inserts inline tags for noise, filler sounds, and garbage, and No-Input interactions can be classified as Correct No-Input or False No-Input to drive Voice Activity Detection tuning.
- Compare and Report — Once transcribed, the portal measures recognition against the ground truth and produces a full set of accuracy statistics: Word Accuracy, Utterance Accuracy, SI Match counts (semantically equivalent recognitions), Out-of-Grammar and Out-of-Coverage rates, and Average Confidence. Per-grammar attribution identifies which grammar handled each match.
- Drill Down — Word, Error, and OOG lists break down individual recognized terms by count, accuracy, average confidence, and average length. Per-interaction views provide playback, download, and a complete metadata inspector (grammars used, configuration applied, raw Engine output).
Filtering and Analysis
Every column in the interaction grid is filterable using a lightweight query syntax. Filters support like, =, and != operators, and prefixing a query with ! makes the match case-sensitive. Double-clicking a column header pre-populates the filter for that field.
Import and Export
Analysis sets can be exported to a portable file and imported into another deployment. Exports may optionally be secured with a passkey at the time of export, which must then be supplied during import. This makes it straightforward to share a curated tuning set between staging and production environments, or to deliver a reference set to a customer.
Related Articles