---
title: "Automatic Speech Recognition for Child Reading: A Phonemic"
description: "Assessing how young children decode written words is usually done by hand: a specialist listens to each child, times the… 25 references."
authors: ["Aline N. Rodrigues", "Carlos H. C. Ribeiro"]
year: 2026
venue: "Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1"
source: "https://aclanthology.org/2026.propor-1.98/"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/rodrigues-2026-automatic-speech-recognition-child-reading/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/rodrigues-2026-automatic-speech-recognition-child-reading/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/23/) / Automatic Speech Recognition for Child Reading: A Phonemic Approach using Isolated Words in Brazilian Portuguese

# Automatic Speech Recognition for Child Reading: A Phonemic Approach using Isolated Words in Brazilian Portuguese

**Authors:** Aline N. Rodrigues and Carlos H. C. Ribeiro · **Year:** 2026 · **Venue:** [Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1](https://md.research.elamur.ai/free/en/venues/propor/) ([International Conference on Computational Processing of Portuguese 2026](https://md.research.elamur.ai/free/en/venues/propor/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.propor-1.98/) · [PDF](https://aclanthology.org/2026.propor-1.98.pdf)

## Summary
Assessing how young children decode written words is usually done by hand: a specialist listens to each child, times the reading, and notes errors. This is slow, subjective and hard to scale. Off-the-shelf ASR handles child speech poorly and returns spelled-out words, which hides the phoneme-level mistakes that educators need to see. The authors address this for Brazilian Portuguese by fine-tuning Wav2Vec2.0 models to output phonemes rather than graphemes. They built a new corpus of isolated words read aloud by 2nd- and 3rd-graders and converted the transcripts to SAMPA with the FalaBrasil G2P tool. A custom tokenizer maps each multi-character SAMPA symbol to a single Unicode token. Training follows a three-phase scheme that adds acoustic augmentation in the middle phase. The best model, a CORAA model fine-tuned with augmentation, cuts the phoneme error rate by roughly three quarters compared with the baseline models. Speech therapists confirmed that many of its apparent errors were in fact faithful transcriptions of the children's real articulation deviations. This makes the system useful as a diagnostic tool for early reading assessment.

## Key points
- **Corpus:** 208 children from Belo Horizonte and São José dos Campos (Southeastern dialect) read a 24-word list (Lúcio et al., 2018).
  - Words were segmented by forced alignment with Whisper large-v2 via stable-ts, giving 5,400 samples (about 2.1 h).
  - The 60/20/20 split is stratified by grade and by reading-score quartile.
- **Base models compared:** BP-adult models (CORAA, Grosman-XLS-R-1B) and multilingual models (XLSR-53, XLS-R-300M/1B).
  - Each model's original tokenizer and output head were replaced, and the last three Transformer layers were reset.
- **Training:** three phases of 30 epochs each: real data, then real plus augmented data, then real data again.
  - The augmented samples were created by adding noise, shifting pitch by up to ±1 semitone, and stretching duration by up to ±5%.
- **Main results:**
  - The best baseline reached WER 0.5410 and PER 0.1771 (Grosman-1b).
  - CORAA-aug reached PER 0.0437, a 75.3% relative reduction (Wilcoxon test, p \< 0.05).
  - Augmentation alone lowered PER from 0.0473 to 0.0437.
  - The multilingual 1B model scored PER 0.0590, so prior Portuguese pre-training beat model size.
- **Further analysis:**
  - PER was 0.041 for 2nd grade and 0.046 for 3rd grade.
  - Cluster Homogeneity of the final-layer embeddings rose from 0.35 to 0.94 after fine-tuning, with phonetically coherent groupings.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

📚 Shelves (2): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works

🧭 Topics (1): [🎙️ Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) · 480 works

## Cite

Cite as: Aline N. Rodrigues and Carlos H. C. Ribeiro. 2026. Automatic Speech Recognition for Child Reading: A Phonemic Approach using Isolated Words in Brazilian Portuguese. Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{rodrigues2026automatic,
  title = {{Automatic Speech Recognition for Child Reading: A Phonemic Approach using Isolated Words in Brazilian Portuguese}},
  author = {Aline N. Rodrigues and Carlos H. C. Ribeiro},
  booktitle = {Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1},
  year = {2026},
  url = {https://aclanthology.org/2026.propor-1.98/}
}
```

- [← previous Automatic scoring of ENEM essays: an empirical study of linguistic and contextual representations](https://md.research.elamur.ai/free/en/papers/matos-2026-avaliacao-automatica-de-redacoes-do/)
- [next → Automatic Speech Recognition for Documenting Endangered Languages: Case Study of Ikema Miyakoan](https://md.research.elamur.ai/free/en/papers/taguchi-2026-automatic-speech-recognition-documenting-endangered/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/23/)

## References

25 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Alair Junio Lemes de Andrade, Letícia Correa Celeste, and Luciana Mendonça Alves. 2019. Caracterização da fluência de leitura em escolares do ensino fundamental ii. Audiology - Communication Research, 24:e1983. [find](https://scholar.google.com/scholar?q=Alair+Junio+Lemes+de+Andrade%2C+Let%C3%ADcia+Correa+Celeste%2C+and+Luciana+Mendon%C3%A7a+Alves%2E+2019%2E+Caracteriza%C3%A7%C3%A3o+da+flu%C3%AAncia+de+leitura+em+escolares+do+ensino+fundamental+ii%2E+Audiology+%2D+Communication+Research%2C)
2. Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong an, Naman Singh, Aditya Saraf, Tatiana Likhomanenko, F-aith Sincl, Alexandre Défossez, and 1 others. 2021. [XLS-R: Self-supervised cross-lingual speech representation learning at scale](https://md.research.elamur.ai/free/en/papers/babu-2021-xls-r-self-supervised-cross/). arXiv preprint arXiv:2111.09296.
3. Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. [wav2vec 2.0: A framework for self-supervised learning of speech representations](https://md.research.elamur.ai/free/en/papers/baevski-2020-wav2vec-2-0-framework-self/). arXiv preprint.
4. Edresson Casanova. 2022. Wav2vec 2.0 trained with coraa portuguese dataset. [https://huggingface.co/Edresson/](https://huggingface.co/Edresson/) wav2vec2-large-xlsr-coraa-portuguese. [↗](https://huggingface.co/Edresson/)
5. Richeng Duan. 2023. Joint learning feature and model adaptation for unsupervised acoustic modelling of child speech. In Proc. [find](https://scholar.google.com/scholar?q=Richeng+Duan%2E+2023%2E+Joint+learning+feature+and+model+adaptation+for+unsupervised+acoustic+modelling+of+child+speech%2E+In+Proc%2E)
6. INTERSPEECH 2023, pages 5227–5231. [find](https://scholar.google.com/scholar?q=INTERSPEECH+2023%2C+pages+5227%E2%80%935231%2E)
7. Jonatas Grosman. 2022. Fine-tuned XLS-R 1B model for speech recognition in Portuguese. [https://huggingface.co/jonatasgrosman/](https://huggingface.co/jonatasgrosman/) wav2vec2-xls-r-1b-portuguese. [↗](https://huggingface.co/jonatasgrosman/)
8. Tamas Grosz, Yaroslav Getman, Ragheb Al-Ghezi, Aku Rouhe, and Mikko Kurimo. 2023. Investigating wav2vec2 context representations and the effects of fine-tuning, a case-study of a finnish model. In Proc. [find](https://scholar.google.com/scholar?q=Tamas+Grosz%2C+Yaroslav+Getman%2C+Ragheb+Al%2DGhezi%2C+Aku+Rouhe%2C+and+Mikko+Kurimo%2E+2023%2E+Investigating+wav2vec2+context+representations+and+the+effects+of+fine%2Dtuning%2C+a+case%2Dstudy+of+a+finnish+model%2E+In+Pro)
9. INTERSPEECH 2023, pages 196–200. [find](https://scholar.google.com/scholar?q=INTERSPEECH+2023%2C+pages+196%E2%80%93200%2E)
10. Jian. 2025. stable-ts: Modifies openai’s whisper to produce more reliable timestamps. [https://pypi.org/project/stable-ts/](https://pypi.org/project/stable-ts/). Version 2.19.0. [↗](https://pypi.org/project/stable-ts/)
11. Alexander Johnson, Hariram Veeramani, Natarajan Balaji Shankar, and Abeer Alwan. 2023. An equitable framework for automatically assessing children’s oral narrative language abilities. In Proc. [find](https://scholar.google.com/scholar?q=Alexander+Johnson%2C+Hariram+Veeramani%2C+Natarajan+Balaji+Shankar%2C+and+Abeer+Alwan%2E+2023%2E+An+equitable+framework+for+automatically+assessing+children%E2%80%99s+oral+narrative+language+abilities%2E+In+Proc%2E)
12. INTERSPEECH 2023, pages 4608–4612. [find](https://scholar.google.com/scholar?q=INTERSPEECH+2023%2C+pages+4608%E2%80%934612%2E)
13. Patrícia Silva Lúcio, Hugo Cogo Moreira, Adriana de Souza Batista Kida, Carolina Alvez Ferreira de Carvalho, Ângela Maria Vieira Pinheiro, Jair de Jesus Mari, and Clara Regina Brandão de Avila. 2018. Word decoding task: Item analysis by irt and within-group norms. Psicologia: Teoria e Pesquisa, 34:e… [find](https://scholar.google.com/scholar?q=Patr%C3%ADcia+Silva+L%C3%BAcio%2C+Hugo+Cogo+Moreira%2C+Adriana+de+Souza+Batista+Kida%2C+Carolina+Alvez+Ferreira+de+Carvalho%2C+%C3%82ngela+Maria+Vieira+Pinheiro%2C+Jair+de+Jesus+Mari%2C+and+Clara+Regina+Brand%C3%A3o+de+Avila%2E+2018%2E+)
14. Bo Molenaar, Cristian Tejedor-Garcia, Catia Cucchiarini, and Helmer Strik. 2023. Automatic assessment of oral reading accuracy for reading diagnostics. In Proc. [find](https://scholar.google.com/scholar?q=Bo+Molenaar%2C+Cristian+Tejedor%2DGarcia%2C+Catia+Cucchiarini%2C+and+Helmer+Strik%2E+2023%2E+Automatic+assessment+of+oral+reading+accuracy+for+reading+diagnostics%2E+In+Proc%2E)
15. INTERSPEECH 2023, pages 5232–5236. [find](https://scholar.google.com/scholar?q=INTERSPEECH+2023%2C+pages+5232%E2%80%935236%2E)
16. Ana Luiza Gomes Pinto Navas, Joana Cecilia Baptista Ramalho Pinto, and Paula Roberta Rocha Del-lisa. 2009. Avanços no conhecimento do processamento da fluência em leitura: da palavra ao texto. Revista da Sociedade Brasileira de Fonoaudiologia, 14(4):553–559. [find](https://scholar.google.com/scholar?q=Ana+Luiza+Gomes+Pinto+Navas%2C+Joana+Cecilia+Baptista+Ramalho+Pinto%2C+and+Paula+Roberta+Rocha+Del%2Dlisa%2E+2009%2E+Avan%C3%A7os+no+conhecimento+do+processamento+da+flu%C3%AAncia+em+leitura%3A+da+palavra+ao+texto%2E+Revista)
17. Ankita Pasad, Ju-Chieh Chou, and Karen Livescu. 2021. Layer-wise analysis of a self-supervised speech representation model. In 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 914–921. [find](https://scholar.google.com/scholar?q=Ankita+Pasad%2C+Ju%2DChieh+Chou%2C+and+Karen+Livescu%2E+2021%2E+Layer%2Dwise+analysis+of+a+self%2Dsupervised+speech+representation+model%2E+In+2021+IEEE+Automatic+Speech+Recognition+and+Understanding+Workshop+%28ASRU%29%2C)
18. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine Mcleavey, and Ilya Sutskever. 2023. [Robust speech recognition via large-scale weak supervision](https://md.research.elamur.ai/free/en/papers/radford-2022-robust-speech-recognition-large-scale/). In Proceedings of the 40th International Conference on Machine Learning, pages 28492–28518.
19. Aline Rodrigues, Gabriela Ribeiro, Victor Silva, Wesley Carvalho, Miguel Ramírez, Luciana Alves, Marcelo Finger, Ana Luiza Navas, and Carlos Ribeiro. 2023. AI and reading fluency for brazilian portuguese: A preliminary study. SSRN Preprint. [find](https://scholar.google.com/scholar?q=Aline+Rodrigues%2C+Gabriela+Ribeiro%2C+Victor+Silva%2C+Wesley+Carvalho%2C+Miguel+Ram%C3%ADrez%2C+Luciana+Alves%2C+Marcelo+Finger%2C+Ana+Luiza+Navas%2C+and+Carlos+Ribeiro%2E+2023%2E+AI+and+reading+fluency+for+brazilian+portugu)
20. Andrew Rosenberg and Julia Hirschberg. 2007. Vmeasure: A conditional entropy-based external cluster evaluation measure. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 410– 420. [find](https://scholar.google.com/scholar?q=Andrew+Rosenberg+and+Julia+Hirschberg%2E+2007%2E+Vmeasure%3A+A+conditional+entropy%2Dbased+external+cluster+evaluation+measure%2E+In+Proceedings+of+the+2007+Joint+Conference+on+Empirical+Methods+in+Natural+Lang)
21. Ana Siravenha, Nelson Neto, Valquíria Macedo, and Aldebaro Klautau. 2008. Uso de regras fonológicas com determinação de vogal tônica para conversão grafema-fone em Português Brasileiro. 7th International Information and Telecommunication Technologies Symposium. [find](https://scholar.google.com/scholar?q=Ana+Siravenha%2C+Nelson+Neto%2C+Valqu%C3%ADria+Macedo%2C+and+Aldebaro+Klautau%2E+2008%2E+Uso+de+regras+fonol%C3%B3gicas+com+determina%C3%A7%C3%A3o+de+vogal+t%C3%B4nica+para+convers%C3%A3o+grafema%2Dfone+em+Portugu%C3%AAs+Brasileiro%2E+7th+Internatio)
22. Anuroop Sriram, Michael Auli, and Alexei Baevski. 2022. Wav2vec-aug: Improved self-supervised training with limited data. Preprint, arXiv:2206.13654. — [arXiv:2206.13654](https://arxiv.org/abs/2206.13654) [find](https://scholar.google.com/scholar?q=Anuroop+Sriram%2C+Michael+Auli%2C+and+Alexei+Baevski%2E+2022%2E+Wav2vec%2Daug%3A+Improved+self%2Dsupervised+training+with+limited+data%2E+Preprint%2C+arXiv%3A2206%2E13654%2E+%E2%80%94+%5BarXiv%3A2206%2E13654%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2206%2E13)
23. Emre Yılmaz, Joris Pelemans, and Hugo Van hamme. 2014. Automatic assessment of children’s reading with the FLaVoR decoding using a phone confusion model. In Interspeech 2014, pages 969–972. [find](https://scholar.google.com/scholar?q=Emre+Y%C4%B1lmaz%2C+Joris+Pelemans%2C+and+Hugo+Van+hamme%2E+2014%2E+Automatic+assessment+of+children%E2%80%99s+reading+with+the+FLaVoR+decoding+using+a+phone+confusion+model%2E+In+Interspeech+2014%2C+pages+969%E2%80%93972%2E)
24. Han Zhu, Li Wang, Gaofeng Cheng, Jindong Wang, Pengyuan Zhang, and Yonghong Yan. 2022. Wav2vec-s: Semi-supervised pre-training for low-resource asr. In Proc. [find](https://scholar.google.com/scholar?q=Han+Zhu%2C+Li+Wang%2C+Gaofeng+Cheng%2C+Jindong+Wang%2C+Pengyuan+Zhang%2C+and+Yonghong+Yan%2E+2022%2E+Wav2vec%2Ds%3A+Semi%2Dsupervised+pre%2Dtraining+for+low%2Dresource+asr%2E+In+Proc%2E)
25. Interspeech 2022, pages 4870– 4874. [find](https://scholar.google.com/scholar?q=Interspeech+2022%2C+pages+4870%E2%80%93+4874%2E)

*Updated 1 November 2026*