Speech recognition in the service
The service uses OpenAI gpt-4o-mini-transcribe for speech transcription. Our Qwen-based ASR remains in research and is not used for service transcription or as a fallback.
Speech AI designed for children’s voices
We build technology that understands children’s voices better and can explain the evidence behind each result.
Internal validation7.05%
Child speech character error rate: 50 sentences from one held-out speaker, with no training overlap
Kaffol AI
From a child’s response to a report
We go beyond turning speech into text. The design considers age in months, the assessment items, and evaluation rules so that the evidence behind report statements can be reviewed.
- 01Child’s response
Child speech and behavior input
Children’s own responses captured through camera, microphone, and touch
- 02Speech recognition
Service speech recognition
The service transcribes children’s speech with OpenAI gpt-4o-mini-transcribe.
- 03Reviewing the evidence
Interpretation linked to evidence
Completed tasks, the child’s actual responses, and scoring criteria for each assessment
- 04Explaining results
A clear, readable report
Describing current observations and next conversations based on actual responses
Internal validation results
Compared using the same child speech samples
We compared the character error rate (CER) of ASR with our own LoRA on the same internal samples. A lower value means a closer match to the original text.
Evaluation conditions50 sentences from one held-out speaker, with no training overlap
This small internal benchmark examines the effect of LoRA on children’s speech. It does not demonstrate superior performance across general use cases.
Character error rateLower is more accurate
| Model | Character error rate |
|---|---|
| Kaffol AI ASR | 7.05% |
| Qwen3-ASR 1.7B base | 8.89% |
| OpenAI gpt-transcribe | 9.56% |
How we describe our technology
We distinguish current work from ongoing research.
A small adapter for children’s speech
We are studying adaptation to children’s speech with an approximately 25 MB LoRA adapter that trains the text decoder while keeping the speech encoder fixed.
Exploring privacy through local AI
As field validation and manufacturing preparation progress, we are exploring local and on-device approaches that process children’s data closer to the device.
Research-based statistical analysis and papers
We are preparing statistical analyses and research papers based on item research. We do not present ongoing research as completed.
Two patent applications relating to local AI
We have filed two patent applications relating to local AI. They are identified as applications, not granted patents.
Research with us