TL;DR
This paper introduces the NITK Kids' Speech Corpus, a novel dataset of Kannada child speech from ages 2.5 to 6 years, addressing the scarcity of resources for this crucial demographic. The corpus enables deeper analysis of developmental speech patterns and supports the development of more robust speech technologies for young children.
What Problem Does This Paper Address?
The vast majority of speech technology applications, from voice assistants to dictation software, are designed and trained on adult speech. While there's a growing recognition of the importance of children's speech for applications like educational tools and games, the availability of child speech datasets is significantly limited, especially compared to adult speech. Existing child speech corpora often focus on older children (6-15 years) and primarily English. This leaves a critical gap for understanding and developing technologies for younger children (2-6 years), a period characterized by rapid developmental changes in speech production due to evolving vocal tract anatomy and neuromotor control. These developmental changes lead to unique speech characteristics, including mispronunciations (phonological processes), which are vital for speech-language pathologists to analyze language development and identify potential disorders. The lack of comprehensive datasets for this age group hinders both fundamental research into child speech development and the creation of effective, age-appropriate speech technologies.
Methods
The paper details the design and collection of the NITK Kids' Speech Corpus. Key aspects include:
- Target Demographic: Children aged 2.5 to 6 years, divided into four age groups (2-3, 3-4, 4-5, 5-6 years).
- Language: Kannada, a South Indian language.
- Recording Protocol:
- Recordings were made in a quiet room using a Blue Yeti USB microphone with a cardioid polar pattern.
- Audio was captured at a 48 kHz sampling rate and 16-bit depth.
- WaveSurfer software was used for recording and annotation.
- To elicit speech from young children who may not be able to read, pictures were used to prompt descriptions of representative words. This approach aims to capture contextualized speech.
- Sufficient breaks were incorporated to maintain child engagement.
- Data Analysis:
- Prosodic Analysis: Pitch variation was analyzed across age groups and genders to observe developmental trends in vocal fold vibration.
- Spectral Analysis: Formant frequencies (F1, F2, F3, F4) for key vowels (/a/, /A/, /i/, /u/, /e/, /o/) were extracted using Linear Predictive Coding (LPC) analysis. This analysis aimed to track changes in vocal tract shape and size with age.
- Quality Control: Recordings were reviewed by phoneticists and Speech-Language Pathologists (SLPs), with rerecording suggested if necessary.
Datasets
- NITK Kids' Speech Corpus (Newly Introduced):
- Language: Kannada
- Age Range: 2.5 to 6 years (divided into 2-3, 3-4, 4-5, 5-6 year groups).
- Participants: 160 children (20 male, 20 female per age group).
- Total Duration: Approximately 10 hours of speech recordings.
- Content: Spontaneous and imitated speech (for 2-3 year olds) and spontaneous speech (for 3-6 year olds).
- Distinguishing Features: Focus on a younger age range than many existing corpora, inclusion of a South Indian language, and a methodology designed for non-reading children.
- Existing Datasets (Referenced in Table 1): The paper references various existing children's speech datasets, primarily in English, for comparison. These include corpora like CID children's speech corpus, CMU Kid's speech corpus, CU Kid's Prompted and Read Speech corpus, OGI Kid's speech corpus, TIDIGITS, and others, mostly targeting older children and read speech.
Key Insights and Ideas
- Developmental Trajectory of Speech: The study highlights that children's speech characteristics, both prosodic (pitch) and spectral (formants), change systematically with age due to physical growth and neuromotor development.
- Gender Differences in Speech Development: The analysis revealed significant differences in pitch between male and female children, with males generally exhibiting lower pitch, attributed to thicker vocal folds.
- Vocal Tract Lengthening: The decrease in formant frequencies with age directly correlates with the lengthening of the vocal tract as children grow, a fundamental aspect of speech development.
- Challenges of Young Child Speech: The paper underscores the inherent variability and complexity in speech production for children under 6, making robust recognition and analysis more challenging than for older children or adults.
- Importance of Contextualized Speech: Using pictures to elicit descriptions is a practical approach to gather meaningful speech data from young children who cannot read, capturing more natural speech patterns.
Key Results
- Pitch Analysis: A systematic decrease in median pitch was observed with increasing age for both male and female children, with consistent gender differences noted across all age groups. For instance, at 3 years old, the median pitch for females was around 380Hz, while for males it was 340Hz.
- Formant Analysis: Scatter plots of F1 vs. F2 for various vowels showed a systematic decrease in formant values from younger to older age groups (e.g., from 3-3.5 years to 6-6.5 years). For vowel /a/, the mean F1 decreased from 975.00Hz to 826.80Hz, and F2 from 2356.10Hz to 2160.50Hz between these age ranges, indicating vocal tract lengthening.
- Corpus Size: The NITK Kids' Speech Corpus comprises approximately 10 hours of recordings from 160 children, providing a substantial resource for its target demographic and language.
Limitations and Future Work
- Limitations:
- The paper acknowledges the difficulty in recording speech from young children due to short attention spans and inability to read, which influenced the data collection methodology.
- While the corpus is designed to capture developmental changes, the analysis presented is preliminary, focusing on pitch and formant trends.
- Future Work:
- The authors suggest the corpus can be used for a wide range of applications, including:
- Developing robust children's speech recognition systems for lower age groups.
- Speaker, age-group, and gender identification from child speech.
- Analyzing language learning ability and detecting mispronunciations (phonological processes).
- Speech-language pathology research.
- Studying the practice of producing various speech sounds.
- Further exploitation of the corpus for detailed analysis of spectral and prosodic features, as well as excitation source characteristics, is envisioned.
Summary updated