> ## Documentation Index
> Fetch the complete documentation index at: https://docs.corti.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Speech-to-Text Performance Evaluation

> Learn how to quantify and visualize speech recognition results

Building with speech-to-text doesn’t stop at transcription. You need to understand how well your system performs in real-world conditions; however, evaluations can be complex with varying methodologies to calculate error and accuracy rates. Understanding what causes degradation in performance is just as important as measuring it in the first place.

<Card>
  This page describes an open source CLI tool offered by Corti, providing a means for consistent, best practice speech to text evaluations.

  <Icon icon="ruler" iconType="solid" />    Measure accuracy beyond simple metrics like Word Error Rate (WER)\
  <Icon icon="briefcase-medical" iconType="solid" />    Identify failures on critical terms (e.g., medical vocabulary)\
  <Icon icon="scale" iconType="solid" />    Compare models, configurations, and prompts reliably\
  <Icon icon="binary" iconType="solid" />    Debug systematic errors and improve product quality
</Card>

***

## A Better Way to Evaluate ASR

Metrics like WER treat all words equally, but in real-world use cases, not all errors carry the same weight. Missing filler words is very different from missing a diagnosis, medication, or key symptom.

**ErrorAlign** is a next-generation alignment and evaluation method designed for modern speech recognition.

Instead of forcing transcripts into rigid, one-to-one comparisons, it produces more natural, human-like mappings between reference and model output. This makes error analysis clearer, more reliable, and more actionable.

<Card title="Click here to read the Error Align paper" href="https://arxiv.org/abs/2509.24478">
      <Icon icon="check" /> Produces more human-like transcript comparisons <br />
      <Icon icon="check" /> Improves error attribution <br />
      <Icon icon="check" /> Enables deeper, more actionable analysis <br />
</Card>

***

## Measure and Visualize Performance

**Corti Canal** is an open source command-line tool that measures and visualizes speech-recognition performance.

Input a CSV with the expected reference transcripts alongside the model-generated output. A self-contained HTML report will be produced containing overall accuracy metrics and a word-by-word comparison, so you can see exactly where the model gets things right and where it does not.

* One to many files can be evaluated in a single run with metrics (e.g., word error rate, character error rate, and medical term recall) calculated across all rows in the CSV.
* Choose between standard Levenshtein distance or Corti’s [Error Align](https://arxiv.org/abs/2509.24478) algorithm for evaluation methodology.
* Medical terms are defined based on your preferences: reference a vocabulary library to be used in the evaluation, augment and maintain the vocabulary over time or customize for a given analysis.
* Analysis normalizes output (all lowercase, without punctuation) by default, but an option to disable normalization is available to include checking of casing and punctuation.
* In addition to the metrics, a visualization of the results is included, overlaying generated and final results to see each error and medical term identified in the results.

<Frame>
  <img src="https://mintcdn.com/corti/ZYFaO_OgxyCo6Otv/images/corti-canal.png?fit=max&auto=format&n=ZYFaO_OgxyCo6Otv&q=85&s=15525218149caf0a95fe1932dd5f5308" alt="Corti Canal" width="1268" height="1047" data-path="images/corti-canal.png" />
</Frame>

<br />

<Card title="Click here to view Corti Canal on Github" href="  https://github.com/corticph/corti-canal">
      <Icon icon="check" /> Built-in normalization <br />
      <Icon icon="check" /> Advanced alignment methods <br />
      <Icon icon="check" /> Visual error analysis  <br />
      <Icon icon="check" /> Domain-specific metrics (e.g., medical term recall)   <br />
</Card>

***

## What This Enables

Instead of merely asking:

> “What is the WER?”

You can answer:

> “Where is my model failing, and does it matter?”

<br />

<Note>
  See some of our supporting research here:

  * [Self-Supervised Speech Representation Learning: A Review](https://ieeexplore.ieee.org/abstract/document/9893562)
  * [Do We Still Need Automatic Speech Recognition for Spoken Language Understanding](https://arxiv.org/abs/2111.14842)
  * [Do End-to-End Speech Recognition Models Care About Context](https://arxiv.org/abs/2102.09928)
  * Learn more at [corti.ai/research](https://corti.ai/research)
</Note>