Even in today’s digitised world, there are still many forms that need to be filled in by hand. Especially forms that need to be filled in in the field. Optical character recognition (OCR) solutions are generally not very good at reading handwritten form entries.

OCR works well on printed text. But it isn’t as good on handwritten information. Also, text printed by a matrix printer is hard to decipher due to inconsistent print quality between forms.


Singer is a service that transcribes handwritten and matrix-printed form entries into machine-readable text objects to be used by other applications. Singer uses artificial intelligence (AI) in combination with advanced image-processing techniques.

Singer uses pre-trained generic AI models and learns from the specific domain business to improve its quality for each specific business domain. Instead of the pre-trained generic models, Singer can also be kick-started based on historic labelled data.

Singer’s learning capabilities are based on feedback techniques and the human-in-the-loop principle.

Singer focuses on short pieces of information, such as codes, dates, numbers and domain-specific words. It is not intended to read entire paragraphs of free text. Singer also detects the presence of specific objects, such as stamps or signatures, and whether fields are filled in or not.

Due to the focus on application-form entries it is extremely well suited as an element in the digitalization chain of automatic form processing. At first, Singer examines the regions of interest (ROI) of the form and applies advanced image-processing techniques on these. This results in a number of small image segments. As a next step, each image segment is sent to the value-detection algorithm, which responds with the interpreted values. In some cases a multistep approach is needed if the algorithm to be used depends on the outcome of a value. The algorithms used and steps are easily configured.

Singer is a learning AI solution and thus it needs to be retrained regularly. New patterns will pass through the system, and through the feedback system, these new patterns and labels will be ingested into the Singer learning algorithm. The feedback system allows clear monitoring of the AI quality.

It is advised that the customer provides a labelled historic data set for training purposes. This way, the solution will be performing at a higher level at the beginning of the project. If this cannot be done, the AI learns from scratch or a general model. In the end, Singer will be equally efficient but it will take longer since it has to learn everything from production activities.


Singer is offered in 2 flavours :

  • Software as a Service (SaaS):
    • The entire Singer solution is hosted and operated by Vectr.Consulting. The customer connects using API endpoints
    • Monitoring is done by the Vectr.Consulting team
  • On-premise:
    • Singer runs on the infrastructure provided by the customer
    • The customer bears responsibility for keeping the system trained

Setting up Singer requires a careful design process that Vectr.Consulting executes, supported by the customer.


Model and algorithm definitions

Initial model trainings and configurations

Go live

MLops: quality monitoring and retrainings

Implementation possibilities

As mentioned above, Singer will be offered to the customer by means of API endpoints, either on prem or via the SaaS offering. The customer is able to choose if he wants to perform the orchestration of the activities themselves, or they can opt to use the Singer out-of-the-box orchestration.

For each implementation it is advised that the customer provides a (labeled) training data set. This way, the solution will be performing at a higher level at the beginning of the project. If this cannot be done, it will start learning from scratch. In the end, Singer will be equally performant but it will take more time since it has to learn everything from production activities.


Singer is well suited to be used in the digitalization of manual document flows. It runs individually or as part of a set of digitalization tools. Individual use can be the scanning and interpretation of:

  • logistical documents (CMR, order sheets, …)
  • quality documents (automotive, health industry, …)
  • banking documents
  • public service docs (travel documents, border control, …)

An example where it is part of a chain of tools can be where the first tool scans the document, then sends it to Singer for analysis. Singer sends the output to a validation tool. Depending on a positive or negative validation, the information a trigger with this information is sent to a CRM or ERP system.

Reference projects

One of the top use cases is reading physician’s handwriting on medical certifications. Vectr.Consulting built the solution for the Christelijke Mutualiteiten. The full project case, with supporting movie clip, can be found here :


Is Singer out of the box available ?

No, there is always an amount of time that Vectr needs to setup the detection algorithm, depending on the number of fields you want to scan.

Does Singer come with pre-trained models ?

Only limited trained models can be offered (date field for example). Since the training is done specifically on the input types, it is not possible to provide other pre-trained models

What is the licensing model ?

Depending on the number of fields you want to scan, the licensing tier is selected. <<insert more info about licensing>>