Audio Augmentation for Non-Native Children's Speech Recognition through Discriminative Learning.

Area of Science:

• Human-computer interaction
• Speech processing
• Second language acquisition

Background:

• Children's increasing interaction with virtual assistants drives advancements in automatic speech recognition (ASR).
• Non-native children exhibit unique speech errors during second language (L2) acquisition, challenging current ASR systems.
• Existing ASR struggles to accurately recognize the speech patterns of non-native children.

Purpose of the Study:

• To develop an effective automatic speech recognition system for non-native children.
• To improve ASR performance by addressing L2 acquisition speech characteristics.
• To investigate the impact of L2 proficiency on ASR systems for children.

Main Methods:

• Utilized feature-space discriminative models, including feature-space maximum mutual information (fMMI) and boosted feature-space maximum mutual information (fbMMI).
• Implemented speed perturbation-based data augmentation on children's speech corpora.
• Collected and analyzed speech data encompassing various speaking styles, including read and spontaneous speech.

Main Results:

• Feature-space MMI models demonstrated superior performance compared to traditional ASR baseline models.
• Increasing speed perturbation factors positively correlated with improved ASR accuracy.
• The developed system showed enhanced recognition capabilities for non-native children's speech.

Conclusions:

• Feature-space discriminative models combined with speed perturbation offer a promising approach for non-native children's ASR.
• Addressing specific challenges in L2 acquisition speech is crucial for robust ASR systems.
• This research contributes to more inclusive and effective human-computer interaction technologies for young learners.