---
title: "Transcription Models Overview"
slug: "transcription-models-overview"
description: "Learn how Hyperscience's transcription model enhances text extraction, automating value acceptance or review based on confidence scores and historical data."
updated: 2026-08-26T13:04:29Z
published: 2026-08-26T13:04:29Z
canonical: "help.hyperscience.ai/transcription-models-overview"
---

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

# Transcription Models Overview

Hyperscience extracts text from documents using Optical Intelligent Character Recognition (OICR). On top of this, the system applies a Transcription model to decide how extracted values are handled — whether they are accepted automatically or sent for human review (e.g.,Transcription Task). To learn more about OICR, see [Text Segmentation](/v43/docs/text-segmentation).

> [!NOTE]
> A Transcription model does not perform text extraction
> 
> It acts as a decision layer that evaluates how reliable extracted values are, based on historical QA data and system configurations.

When text is extracted, not every value is equally reliable. Some predictions are clearly correct, while others are uncertain or incorrect.

Without a transcription model, the system relies on fixed confidence thresholds to decide:

- accept the value automatically
- or send it for manual review.

> [!NOTE]
> Transcription models change how predictions are evaluated
> 
> - If no transcription model is configured, the system uses the Threshold set in the flow.
> - If a transcription model is configured, the manually configured threshold is overridden. The model uses QA performance and target accuracy to determine the applicable threshold.

In this article, you’ll learn:

- What a transcription model is and what it does.
- How the system decides whether to accept a value or send it for review.
- The difference between thresholds and transcription models.
- How Target Accuracy controls the balance between Automation and manual review.

> [!NOTE]
> Manage a Transcription model
> 
> For instructions to view, assign, train, monitor, download, import, deploy, or deactivate a model, see [TDM for Transcription Models](/v43/docs/tdm-transcription-models).

## Understanding Transcription models

A transcription model is a machine-learning model that helps the system adapt how it decides whether to accept extracted values automatically or send them for review.

When text is extracted, the system assigns a confidence score to each value. Instead of relying only on static thresholds, the transcription model uses QA feedback to learn how reliable those predictions are over time.

This learned behavior is then used to decide whether a value should be:

- accepted automatically, or
- sent for human review.

> [!NOTE]
> Transcription models and flows
> 
> - Each Document Processing flow can be associated with a transcription model.
> - A single transcription model can be reused across multiple flows, depending on how it is configured.
> - Within a flow, the assigned transcription model controls how extracted values are evaluated and whether they are accepted automatically or sent for review.

- By default, the system includes a transcription model named IDP.
  - This model is trained on transcription QA data collected across all Workflows, in which extracted values are reviewed and corrected during QA tasks..

![](https://cdn.us.document360.io/87894cef-4958-4f3f-be6f-b75a78c82548/Images/Documentation/image(254).png?sv=2026-02-06&spr=https&st=2026-08-21T21%3A25%3A42Z&se=2026-08-21T21%3A40%3A42Z&sr=c&sp=r&sig=%2BI6QV1bSb1qAUP4tPUngPiGUUG7li3Edye2fu1W4oh0%3D)

### Process

For each extracted value, the system determines whether it should be accepted automatically or sent for human review.

1. The system extracts text using OICR.

2. A confidence score (between 0 and 1) is assigned to the extracted value.

> [!NOTE]
> Confidence score
> 
> A confidence score is an internal model score between 0 and 1 that the system uses to compare predictions. It is not an accuracy percentage or a direct probability that a value is correct.
> 
> - Higher values indicate higher confidence.
> - Lower values indicate more uncertainty.

3. The system evaluates the prediction based on the configuration of your flow:

- **If no transcription model is configured:**
  - The system uses the confidence threshold set in the flow.
    - Values with a confidence score above the threshold are accepted automatically.
    - Values below the threshold are sent for review.
- **If a transcription model is configured:**
  - The system no longer uses the manually configured threshold (even if it is still visible in the UI).
  - Instead, it determines the applicable accuracy threshold from:
    - the model’s learned performance (based on QA data), and
    - the target accuracy configured in the flow.

4. Based on this evaluation, the system decides whether to:

- accept the value automatically, or
- send it for review.

> [!NOTE]
> The confidence score is used internally to compare predictions.
> 
> It is not directly equivalent to a percentage and should not be interpreted as an exact probability.

#### Thresholds

If no transcription model is used, the system relies on **static thresholds**.

A threshold defines the minimum confidence required for a value to be accepted automatically.

- **Example 1:**
  - Confidence score: 0.82
  - Threshold: 0.80 → value is accepted.
- **Example 2:**
  - Confidence score: 0.82
  - Threshold: 0.90 → value is sent for review.

This approach is fixed and does not adapt over time.

> [!NOTE]
> When a transcription model is deployed:
> 
> - The manually configured static threshold is no longer used to make decisions (even if it is still visible in the UI).
> - The system instead applies an accuracy threshold derived from the model’s learned behavior and the configured target accuracy.
> 
> In practice, **the adapted model replaces the manually configured threshold with a learned threshold** that learns based on past performance.

#### Target accuracy

Target accuracy defines how many errors are acceptable for your use case.

- For example:
  - 95% target accuracy allows up to 5 incorrect values per 100 fields.
  - 99% target accuracy allows up to 1 incorrect value per 100 fields.

This setting controls how strict the system is when deciding whether to accept values automatically or send them for review.

- Lower target accuracy → higher automation, more errors.
- Higher target accuracy → more review, fewer errors.

> [!NOTE]
> Target accuracy is not a per-value guarantee
> 
> Target accuracy describes the acceptable error rate across an evaluated population. It does not guarantee that every automated value is correct.
> 
> - For example, a 99% target accuracy can still permit an incorrect value, and a high machine-confidence score does not change this limitation.
>   - If every occurrence of a field requires review regardless of machine confidence, configure Transcription Supervision in the layout or in the flow for that field or use case.
>   - No confidence threshold, target-accuracy value, or Transcription model guarantees that every machine transcription will be correct.

#### Threshold scope and field-level targets

By default, the model uses the same rules for all fields it handles. It does not learn separate rules for individual fields.

You can use Field-Level Accuracy Targets (FLAT) to set a higher or lower target accuracy for selected fields or table columns. A field-level target overrides the flow-level target for that entry and makes routing stricter or less strict. It does not create a separate field-specific model. For example, you can set a higher target accuracy for an account number or email address than for less critical fields.

#### Improved Threshold Accuracy

The **Improved Threshold Accuracy** flow setting helps the system meet the target accuracy for both blank and non-blank fields instead of evaluating only their combined accuracy. This distinction matters because OICR is generally better at recognizing blank fields than transcribing populated fields.

For example, suppose the system processes 1000 fields:

- 500 fields are blank, and all 500 are handled correctly.
- 500 fields are populated, and 490 are transcribed correctly.

The combined accuracy is 99%, but the accuracy for populated fields is only 98%. When **Improved Threshold Accuracy** is enabled, the system selects the threshold so that the target is evaluated for the blank and non-blank groups separately. This prevents strong performance on blank fields from masking lower performance on populated fields.

### QA data and transcription models

The Transcription model learns from Quality Assurance (QA) records that indicate whether machine predictions were correct or incorrect. QA data does not change a live model immediately. It affects behavior only after a training cycle updates the applicable model.

High-quality QA data is more important than raw volume. Training Data should:

- Represent the documents and field values processed in production.
- Include realistic blank and non-blank fields.
- Include both correct and incorrect machine predictions. Incorrect examples are especially valuable because they show the model where predictions should not be trusted; their absence does not prevent learning, but it can limit how effectively the model adjusts to errors.
- Contain the correct value after any applicable Normalization or formatting.
- Be reviewed for systematic keying, normalization, or Consensus errors.

Incorrect or unrepresentative QA records can reduce projected or actual automation and can cause the model to make poorer routing decisions. Large volumes of QA data do not compensate for poor data quality or a distribution that does not match production.

For QA data to affect model behavior, sufficient eligible records must be available, and training must be enabled in the flow configuration or started manually from **Models** > **Transcription**.

> [!NOTE]
> The system trains on a subset of recent QA data
> 
> The system uses up to 50,000 recent QA records when training a Transcription model. If more records are available, only the most recent eligible subset within the configured training period is used. Changes in that period or the composition of the subset can therefore change model behavior.

#### QA data requirements

The minimum number of QA records depends on the document and field type:

- **Structured text fields:** 5000 records.
- **Semi-structured text fields:** 2000 records.
- **Table cells:** 2000 records.
- **Checkboxes:** 2000 records.
- **Signatures:** 2000 records.

Until the required number of records is available for a model type, the system uses the manually configured confidence threshold for that type. More data does not always produce better behavior; quality and representation remain the primary considerations.

## Training a transcription model

Training in the context of transcription models is the process through which the system uses QA data to improve how it evaluates extracted values. As described above, the system learns from QA records over time.

During training, it:

- analyzes which predictions were correct
- identifies where mistakes occurred
- adjusts how confident it should be in similar cases.

This results in more informed decisions about when to trust extracted values and when to require human review.

Training does not happen automatically after every QA action. For training to start:

- Sufficient QA data must be available.
- Training must be enabled in the flow configuration (**Transcription Automation Training**). To learn more, see [Flow Blocks](/v43/docs/flow-blocks#transcription-blocks).

![](https://cdn.us.document360.io/87894cef-4958-4f3f-be6f-b75a78c82548/Images/Documentation/transcription_automation_structured(1).jpg?sv=2026-02-06&spr=https&st=2026-08-21T21%3A25%3A42Z&se=2026-08-21T21%3A40%3A42Z&sr=c&sp=r&sig=%2BI6QV1bSb1qAUP4tPUngPiGUUG7li3Edye2fu1W4oh0%3D)

Once these conditions are met, the system periodically updates model behavior based on the available QA data.

> [!NOTE]
> Daily autotraining enabled
> 
> - If you enable the **Daily autotraining enabled** setting, the system trains new model versions daily from the latest eligible QA records.
> - If you disable it, the deployed model remains live, but new QA data does not update it through scheduled training. You can still start training manually from **Models** > **Transcription**.
>   - Disabling daily autotraining can stabilize model behavior while you investigate a regression, but it also prevents the model from adapting automatically to representative new data.
> 
> ![](https://cdn.us.document360.io/87894cef-4958-4f3f-be6f-b75a78c82548/Images/Documentation/daily_autotraining.jpg)

### Model behavior

The system compares the model’s learned performance (based on QA data) and the configured target accuracy.

Based on this, you may observe:

- **High automation (few or no review tasks):**
  - when target accuracy is lower than model performance.
- **High review volume:**
  - when target accuracy is very strict.
- **Balanced behavior:**
  - when target accuracy aligns with model performance.

For example, if a flow is configured with **Target accuracy: 95%**, over time, based on QA data, the system will learn that its predictions are generally reliable for this use case. As a result:

- The system accepts most values automatically.
- Few values are sent for review.
  - This can lead to high automation, even if some incorrect values are still present.

> [!NOTE]
> Transcription models do not affect how text is extracted
> 
> They do not improve how the system reads the document. Instead, they influence how confident the system is in extracted values and how those values are handled.

## Respond to unexpected model behavior

If automation or error patterns change after training:

1. Identify the affected flow, layout, field type, and model version.
2. Compare Projected Automation with actual automation.
3. Review recent Transcription QA records for incorrect, inconsistent, or unrepresentative values.
4. Check whether the training period, layouts, normalization behavior, or data-retention settings changed.
5. Review the Transcription-task queue to determine the operational impact.
6. If necessary, deploy the previous compatible model version while you investigate.
7. Consider disabling daily autotraining temporarily if new nightly versions would make the investigation harder.
8. After correcting the data or configuration, run training manually and monitor the new version before returning to normal operations.

Switching from an adapted model to manually configured thresholds may reduce automation and does not guarantee higher accuracy. Evaluate that change with representative production data over an appropriate period instead of assuming that one routing method is safer.

For instructions to monitor reports, run training manually, or deploy a previous model version, see [TDM for Transcription Models](/v43/docs/tdm-transcription-models).

## Recalibration

Recalibration updates how the system evaluates extracted values after changes to the underlying models. This typically happens after a platform upgrade, when internal model behavior may change. Recalibration helps the system stay aligned with changes in underlying models, ensuring that past QA data remains useful whenever possible.

> [!NOTE]
> After upgrades, you may notice changes in the behavior
> 
> Recalibration helps align previous QA data with the upgraded system.

When the system is upgraded:

- predictions may change
- previously collected QA data may no longer reflect current behavior.

Recalibration ensures that the system:

- re-evaluates existing QA records using the updated models
- aligns its behavior with the latest model performance.

During recalibration, the system:

- reprocesses available QA records
- compares new predictions with the recorded QA values
- updates how it evaluates prediction reliability.

This allows the model to continue learning from existing data instead of relying only on new QA inputs.

> [!NOTE]
> Recalibration may not be possible if:
> 
> - required crop data is missing.
> - data retention policies have removed necessary inputs. Learn more in [PII Data Deletion](/v43/docs/pii-data-deletion).
> - In these cases:
>   - previously collected QA data cannot be reused.
>   - QA collection may need to start again.

### Requirements

Recalibration depends on:

- Training components (Trainer).
- Recalibration requires the appropriate training setup to be available. If it is not configured, recalibration cannot be performed. Contact your Hyperscience representative for more information.
- Availability of crops:
  - The system must be able to reprocess the original document regions. If this data is missing, recalibration may not be possible.

> [!NOTE]
> Recalibration does not always start from scratch
> 
> - Existing QA data may still be reused, depending on availability.
> - In some cases, additional QA work may be required.

## Troubleshooting and common scenarios

This section addresses common situations observed in real-world usage and explains why they occur. If the system behaves differently than expected, it is usually due to the relationship between target accuracy, QA data, and recent system behavior.

### A high-confidence value is incorrect

A value can be accepted automatically even when the extracted text is incorrect.

#### Why this happens:

- Machine confidence is not the probability that a value is correct.
- A target below 100% explicitly permits errors, and no target guarantees that every result is correct.
- QA data may not contain enough representative examples of the pattern.
- The prediction can be an outlier even when aggregate performance meets the target.

#### What to do:

- Review the image, OICR output, machine-confidence score, model version, and applicable target accuracy.
- Check recent QA records for similar values.
- Set a stricter field-level target for a high-risk field.
- Configure Transcription Supervision when every occurrence requires human review.

### No values are sent for review

You may observe that all values are accepted automatically, even when some incorrect extractions are present.

#### Why this happens:

- Target accuracy is set too low.
- The system is allowed to accept more errors.
- Based on QA data, the system determines that most predictions fall within the allowed error tolerance.

#### What to do:

- Increase the target accuracy.
- Monitor how many values are sent for review.
- Adjust until you reach the desired balance between automation and accuracy.

### High automation but incorrect results

It is possible to observe 100% automation while still having incorrect values.

#### Why this happens:

- The system is configured to prioritize automation.
- Target accuracy allows a higher error rate.
- The system is operating within the defined tolerance.

#### What to do:

- Increase the target accuracy.
- Review QA data to ensure it reflects correct values.

### No improvement after QA

You may perform QA tasks but not see any change in system behavior.

#### Why this happens:

- Training has not been applied, or daily autotraining is disabled and training has not been started manually.
- There are no incorrect examples in the QA data.
- QA data is not representative of real-world documents.

#### What to do:

- Ensure the minimum number of eligible QA records is available, then enable training in the flow or start training manually.
- Include both correct and incorrect examples in QA.
- Verify that QA reflects actual document content.

### Threshold values appear incorrect or unusually low (for example, 0.03)

You may notice that threshold values in the UI appear unexpectedly low or inconsistent.

#### Why this happens:

- The manually configured threshold is **not** used when a model is deployed. The model applies a learned accuracy threshold based on QA performance and target accuracy.
- The value shown reflects the last manually configured threshold.
- It does not affect current system behavior.

#### What to do:

- Do not rely on the manually configured threshold when a model is active.
- Review the target accuracy, QA data, and projected automation instead.

### Model behavior changes unexpectedly

You may observe changes in automation or review patterns over time.

#### Why this happens:

- New, incorrect, or unrepresentative QA data has influenced model behavior.
- Recent data differs from earlier patterns.
- System updates or recalibration have changed how predictions are evaluated.

#### What to do:

- Review recent QA data for consistency.
- Compare projected and actual automation, and monitor Transcription-task volume.
- If necessary, deploy an older compatible model version and pause daily autotraining while you investigate.

### Training does not produce expected results

In some cases, training may not lead to improvements.

#### Why this happens:

- QA data is limited or not representative.
- The system uses a subset of recent QA data.
- There is insufficient variation in the dataset.

#### What to do:

- Ensure QA data covers realistic document scenarios.
- Include edge cases and error scenarios.
- Continue collecting QA data over time.

A technology that automatically identifies and converts printed or handwritten text within digital images into machine-encoded text.

In Hyperscience, OICR is a key step that enables the platform to extract and work with text during document processing.

A Supervision task that allows you to review or enter text the system couldn’t confidently read from a document. This task enables you to ensure accurate final data when the system’s confidence is low.

**Quality Assurance (QA)** is a process that ensures the accuracy and reliability of system outputs. In Hyperscience, QA tasks allow users to review and correct errors in classification, identification, VLM extraction and transcription. Documents may enter QA after processing through Submissions, depending on the workflow configuration.

The confidence limit used to decide if a machine prediction should be sent for human review to ensure accuracy.

A setting specified by the user. It indicates the desired overall system accuracy, including tasks performed by humans. It allows you to evaluate how well the system is expected to perform. See also Field-Level Accuracy Targets (FLAT).

**Automation** refers to the processing of data without the need for human intervention.

A **Workflow** or **Flow** is a customizable workflow in Hyperscience that automates the processing of documents, including steps like classification, data extraction, validation, and output. Flows streamline operations by handling tasks step-by-step with minimal manual effort. Learn more in [Flows](https://help.hyperscience.ai/latest/docs/flows).

A configuration in Hyperscience that allows you to set different accuracy levels for specific fields or table columns. For example, if you need higher accuracy for fields like addresses or account numbers, you can set a higher target for them while keeping other fields at a lower target accuracy. Doing so helps improve the precision of critical fields without adding extra tasks.

Process that ensures the accuracy and reliability of system outputs. In Hyperscience, QA tasks allow users to review and correct errors in classification, identification, VLM extraction and transcription. Documents may be randomly sampled for QA from all processed data.

The input used to teach machine learning models how to process documents accurately. Its structure depends on the model type:

- For **Classification models**, training data consists of uploaded document pages grouped by layout.
- For **Identification models**, training data includes manually annotated fields and tables to train the model to extract specific data points.

The process of converting extracted data into a consistent format. In Hyperscience, normalization helps standardize values like dates, amounts, or addresses so they’re easier to use in downstream systems.

A process used to confirm the correct value of a transcribed field. Consensus is reached when two matching transcriptions are provided for the same field, usually one from a human and one from the machine, or two separate human-provided entries. This process ensures higher accuracy, especially when the system's confidence is low.

The predicted automation based on the desired target accuracy. The projection is derived from the model’s training data. The system automatically ensures that the same data is not used for both projections and training.

An image of the specific field you want to extract.

A separate machine dedicated to handling resource-heavy tasks like training Identification models. It operates independently and connects to the main application through the API.

## Related

- [Building a Structured Use Case](/building-a-structured-use-case.md)
- [Managing Transcription Models](/managing-transcription-models.md)
- [Transcription Settings](/transcription-settings.md)
- [Transcription Quality Assurance](/transcription-quality-assurance.md)
- [Transcription Accuracy and Automation](/transcription-accuracy-and-automation.md)