TL;DR
This paper introduces two novel Russian child speech databases, "INFANTRU" and "CHILDRU," the first of their kind, to support research into language acquisition and automated speech recognition for children. These comprehensive corpora capture longitudinal vocalizations and speech from infancy through early childhood, addressing a critical gap in existing resources.
What Problem Does This Paper Address?
The development of robust automatic speech recognition (ASR) systems for children is significantly hampered by a lack of diverse and comprehensive speech datasets. While adult ASR has benefited from large-scale databases like SpeechDat, similar resources for children, especially for languages other than English, are scarce. This scarcity is particularly acute for early vocalizations and speech development in the first few years of life. Existing child speech databases often focus on specific age groups or tasks (e.g., reading) and lack the longitudinal and varied contextual data needed to understand the nuances of language acquisition and to train reliable child-specific ASR models. This paper aims to fill this gap by creating and describing two extensive Russian child speech corpora.
Methods
The authors describe the meticulous process of data collection for two distinct corpora:
- INFANTRU: This corpus focuses on infants and toddlers from 3 to 36 months old. Data collection was longitudinal, with recordings made every three months for 76 children. It includes spontaneous vocalizations and speech, as well as mother-child interactions in various situations (face-to-face, play, reading, spontaneous interaction). A unique aspect is the inclusion of data from 88 orphans, recorded under less controlled conditions.
- CHILDRU: This corpus targets older children, aged 4 to 7 years. It contains a larger volume of speech material, including spontaneous speech, answers to questions, reading, poetry, counting, alphabet recitation, and play. The data was collected in homes, kindergartens, and child homes.
Both corpora utilize standard audio recording equipment (Marantz PMD222 recorder with a Sennheiser e835S microphone and Marantz PMD660 digital recorder) and are processed using Cool Edit Pro and Praat. The speech files are in Windows PCM format at 22050 Hz, 16-bit. The paper also details a structured naming convention for speech files to ensure easy identification and retrieval based on child, age, and situation. A custom database management program (VDB.EXE) was developed to facilitate searching and playback of recordings based on various child and recording characteristics.
Datasets
The paper introduces two newly created datasets:
- INFANTRU: Contains 2967 recordings (70 hours) from 99 children (3-36 months old), including 16 children with neurological disorders. Longitudinal data from 76 children is a key feature.
- CHILDRU: Contains 28079 recordings (20 hours) from 150 children (4-7 years old), including 142 children from families and 8 orphans.
These datasets are designed to be comprehensive, covering a wide age range and diverse speech situations, and are intended to serve as benchmarks for future research in Russian child speech.
Key Insights and Ideas
- The critical need for longitudinal data: The INFANTRU corpus's longitudinal design is crucial for tracking developmental changes in vocalizations and speech from infancy. This allows researchers to observe the gradual emergence of speech-like sounds, phonemes, and words.
- Variability in child speech development: The paper highlights differences in speech development between normally developing children and those with neurological disorders, particularly in the timing of first words and the prevalence of different sound categories.
- Context matters for speech analysis: The authors emphasize that the context of speech production (e.g., spontaneous speech vs. answering questions) significantly influences articulation and the complexity of language used by children. For instance, children tend to make more articulation mistakes when answering adult questions, possibly because adults stimulate more complex language use.
- Consonant substitution patterns evolve: The analysis of consonant substitutions reveals age-specific patterns, with fricative substitutions being common in younger children and a reduction in the variety of mistakes observed as children age.
- Vowel acoustics show developmental trends: The acoustic characteristics of vowels, as visualized through formant triangles, demonstrate differences between children of different ages and adults, indicating ongoing development in vocal tract control.
Key Results
- INFANTRU: Vowel-like sounds are prevalent in normally developing and at-risk infants during the first year. At 12 months, normally developing children have a higher proportion of syllables compared to consonants and vowels than children in the risk group. First words appear around 12 months for normally developing children, while they are absent in the risk group up to 24 months.
- CHILDRU: Two-syllable words are prevalent in the lexicon of 4-7 year olds, with an increase in words containing three or more syllables as age progresses. Children tend to make more articulation mistakes when responding to adult questions compared to spontaneous speech.
- Acoustic analysis: Formant triangles for vowels show distinct shapes and orientations for different age groups and contexts, reflecting developmental changes in speech production.
Limitations and Future Work
- Limitations: The paper acknowledges that collecting data from orphans was more challenging, leading to less varied recording situations compared to children from families. The description of the "CHILDRU" corpus's speech file format is less detailed than for "INFANTRU."
- Future Work: The authors suggest that these corpora can be used for scientific projects on Russian language acquisition and for developing automated child speech recognition systems. Further analysis of the collected data is implied for deeper insights into speech ontogenesis.
Summary updated