# Reading Adventure — AI feasibility and speech evaluation

8 September 2026 · Technical proposal accompanying the product plan · No speech service has been integrated or tested; no children's audio has been collected.

This English presentation uses the name **Reading Adventure** and the mascot **Fox**. The intended audience remains Dutch-speaking children aged approximately 6–9 in Flanders. An English demonstration is a presentation adaptation; it does not establish that the learning content or speech technology works for that target audience.

## Recommendation

**Build a complete practice experience first, and evaluate automatic assessment of reading aloud in a separate study.** Dutch speech recognition and Flemish voices for reading aloud are available. Their availability does not establish reliable reading feedback for children aged 6–9 in Flanders. A transcript may contain the intended word even when a child hesitated, changed a sound or corrected themselves.

This recommendation follows the product plan and the work shared by the learning-design and prototype agents: around ten sessions reviewed for educational suitability would form the initial exercise library. The interactive prototype records attempts, help and skipped activities. Only a fixed comprehension question receives feedback on the selected answer. Articulation exercises remain a separate, later development track.

## Candidates and limitations

| Component | Documented capability | Decision for Reading Adventure |
|---|---|---|
| Speech to text | Azure Speech lists `nl-BE` for speech recognition. | First candidate for a focused evaluation using Flemish children's voices. This is not yet a supplier selection or a quality claim. |
| Reviewed examples read aloud | Azure TTS offers voices including `nl-BE-DenaNeural` and `nl-BE-ArnaudNeural`. | A possible way to produce fixed example words and sentences in advance, for review by a Flemish specialist. Individual letter sounds should be recorded or reviewed separately. |
| Pronunciation assessment | Azure lists Dutch (`nl-NL`), but not Belgian Dutch (`nl-BE`), for pronunciation assessment. | Do not use it directly as a standard for Flemish pronunciation. First investigate which pronunciation variants and errors it distinguishes. |

The language and voice listings come from the [official Azure language table](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support). Azure also explains that pronunciation assessment uses a different model from ordinary transcription. Comparing text therefore does not replace speech-sound analysis. [Azure — pronunciation assessment](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-pronunciation-assessment).

A second supplier becomes useful if the first candidate falls short on the evaluation set. Google lists `nl-BE` in Speech-to-Text V2, but in the EU table reviewed it appears under the `telephony` and `telephony_short` models; `nl-NL` has different model options. A language code alone therefore does not establish suitability for a child reading on a tablet. [Google — supported languages and models](https://docs.cloud.google.com/speech-to-text/docs/speech-to-text-supported-languages).

Research by Gao and colleagues tested several recognisers on Dutch read-aloud speech from 71 children aged 6–13. The best word transcription did not mean the best detection of reading errors: models could omit or alter incorrect or partial attempts. This supports evaluating feedback separately. The study does not validate this app, current cloud services or children aged 6–9 in Flanders. [Gao et al., Interspeech 2024](https://www.isca-archive.org/interspeech_2024/gao24d_interspeech.pdf).

## What the demonstration can actually show

- One short session with large text, words, a sentence, help and a comprehension question.
- A recap showing activities offered or viewed, activities marked as attempted, help and skipped items. A tap is not evidence of correct reading. Distinguish a request for help from audio actually played; completed playback still does not prove that the child listened.
- Optional spoken examples using an available local browser voice. For the intended Dutch-language pilot, select only voices with `localService === true` and a Dutch language code, preferring `nl-BE`. The English presentation should apply the equivalent check for a local English voice. If no suitable voice is available, show helper text. Do not automatically switch to a remote voice or a different language.
- At this prototype stage, no microphone recording, speech score, automatic reading-level assessment or persistent child progress.

The Web Speech specification distinguishes local and remote voices through `localService`. Voice availability and pronunciation must be checked on the actual target devices. The read-aloud feature does not justify a general claim that a hosted website processes no data at all. [Web Speech API — voice attributes](https://webaudio.github.io/web-speech-api/#dom-speechsynthesisvoice-localservice).

AI could later prepare exercise variations within defined constraints. A specialist should review vocabulary, permitted letters, sentence structure and content before publication. Simple rules can select the next exercise; an open conversation with a language model is not necessary for this first version.

## Practical test matrix for the later speech evaluation

This is a proposed evaluation, not a completed one. Use only lawfully obtained audio authorised for this purpose. Run the initial evaluation without delivering automatic corrections to children. Two reviewers should assess the same recordings independently and discuss disagreements. A Flemish speech and language therapist or reading specialist should define acceptable language variants, reading errors, self-corrections and unclear recordings in advance.

| Group or situation | Include in the evaluation set | What matters |
|---|---|---|
| Age and starting level | Children aged 6, 7, 8 and 9, with different letters already taught and different reading levels | Results by child and level, so averages do not conceal poor performance |
| Language variation | Different Flemish accents and children with other home languages | Do not treat acceptable variants as reading or pronunciation errors; report uncertainty by group |
| Sounding out | Letter-by-letter reading, individual sounds, segmenting and then blending | Preserve all attempts instead of assessing only the final complete word |
| Reading strategies | Hesitations, repetitions, rereading and self-correction | Distinguish a strategy from an error that remains after the attempt |
| Word errors | Omissions, additions and substitutions, including a plausible alternative word | Compare both the error and its location with human assessment |
| Displayed text as context | The same audio with and without additional text context | Check whether context silently changes errors into the intended text |
| Help | First attempt, listening, reading together and trying again | Do not count performance after help as independent mastery |
| Sound conditions | Quiet speech, whispering, silence, background speech and an overlapping parent's voice | Do not score unclear audio or other voices as the child's performance |
| Devices | Intended tablets, phones and computers, browsers and microphones; vary distance | Measure errors and waiting time for each real combination |
| Failures | Microphone permission denied, connection loss, time-outs, double taps and stopping | The microphone stops; no invented score or automatic level reduction; the activity remains skippable |
| Example audio | Every word and sentence at normal and slower speeds | Expert review of sounds, stress, intelligibility and meaning |
| Frustration and support | Observed sessions with an adult available | Required parental help, repeated false corrections, disengagement and the ability to recover |

Keep children in the development set separate from those in the final evaluation. Include new words and sentences so that remembering activities is not confused with generalisation. Record the model version, language setting, text context and decision rules.

Alongside transcription word errors, measure **false corrections**, **missed reading errors** and **the proportion of uncertain outcomes**. Compare transcription with a verbatim human transcript, not only with the exercise text. Report counts, definitions and uncertainty intervals by child, session and relevant group. Supplier scores are not automatically calibrated probabilities.

Agree concrete acceptance thresholds with the specialists before the final evaluation. Keep automatic feedback disabled until sufficient evidence is available and the thresholds are met. A usability pilot with ten to twenty families can reveal practical problems; that sample size alone does not establish reliability or learning gains.

## Minimal data flow for a later version

Proposed flow: **parent setup → short, intentional recording → the app's secure server → selected speech service → limited exercise result → child and parent**. The exercise library and example audio can be prepared in advance. API keys remain on the server. Start recording only for an answer, and always provide a clear stop button.

Do not send a name, date of birth or diagnosis with the audio. Process the recording temporarily; do not retain raw audio or full transcripts by default. Save only necessary progress data, such as the exercise, attempt, help and uncertainty, under a pseudonym. Prevent audio and transcripts from ending up in error logs, analytics or backups. Research recordings require a separate choice, an access list and a retention period.

Microsoft describes processing real-time audio in server memory without storage at rest, and states that it does not retain data for services including real-time recognition and pronunciation assessment. Before implementation, check the chosen service, region, logging, contracts and the app's own server settings; this product documentation does not automatically make the entire app privacy-compliant. [Azure — data processing](https://learn.microsoft.com/en-us/azure/ai-foundry/responsible-ai/speech-service/speech-to-text/data-privacy-security?view=foundry-classic).

When a service offered directly to children relies on consent, Belgium requires consent from the person with parental responsibility for children under 13. The lawful basis and data processing must be established for the specific service. [Belgian Data Protection Authority — consent](https://www.gegevensbeschermingsautoriteit.be/professioneel/avg/rechtsgronden/toestemming).

## Honest pitch copy

**Now:** “Discover Reading Adventure: a concept for short, playful reading activities. Try words, listen to examples and answer a question about the sentence.”

**Ambition:** “We want to use AI to make practice better suited to each child and support reading aloud. We are exploring this with reading specialists and families.”

**Beside the demo:** “This demo does not listen to or assess your voice. The recap shows what you tried and which help you used.”

Claims of proven reading gains, flawless pronunciation assessment, fully independent learning or replacement of speech and language therapy are not yet supported. Assess affordability later using measured audio minutes, hosting, educational review and support costs; do not make an untested price promise now.