1　Introduction

This paper discusses how the acoustic differences in vowel quality primarily depend on the frequency distribution of the first three formants, while traditional Linear Prediction Coefficient (LPC)-based formant extraction algorithms have significant limitations: they tend to produce "spurious formants" for front high vowels (such as [i]) and "formant merging" for back high vowels (such as [u]), and the LPC order requires manual setting without adaptive adjustment, resulting in tedious and inefficient analysis of diphthong syllables. To address these issues, an automated extraction model is proposed, which uses neural networks to automatically identify speech types (voiceless segments, front high vowel segments, back high vowel segments, and other voiced segments) on a frame-by-frame basis, thereby adaptively selecting different LPC orders for formant extraction. This model is implemented directly on the Praat phonetics experimental platform rather than the Pytorch framework, to overcome the disadvantages of existing neural network solutions such as high programming requirements, complex environment configuration, and high hardware costs, ultimately achieving the goals of low cost, user-friendliness, and easy reproducibility. Finally, using specific word examples from Beijing Mandarin and the Huilai dialect point of Chaoshan dialect, the effectiveness of this method in accurately extracting formants of front and back high vowels and presenting dynamic acoustic characteristics within syllables is verified.

2　Principles of Neural Network Models

This section introduces the basic principles of neural network models and their application process in speech type recognition. Neural networks can model complex mapping relationships without analytical solutions through nonlinear operations and iteratively trainable parameters. Their multi-layer structure, based on matrix multiplication with nonlinear functions connecting layers, enables them to fit arbitrarily complex relationships. For speech frame classification tasks, the input consists of acoustic features such as LPC or MFCC, and the output is speech type labels, requiring optimization methods such as gradient descent or Adam algorithm for parameter training. Model development includes two core steps: training and testing, with datasets divided into training and test sets—the former for parameter optimization and the latter for evaluating generalization capability, with sufficient and diverse data being the key prerequisite for ensuring model performance.

3　Data Sources and Preprocessing of Chinese Dialect Audio

This section introduces the collection and preprocessing methods of Chinese dialect audio data. The data comes from single-character audio recordings of 12 speakers from 10 dialect points, using the "Chinese Language Resource Survey Manual·Chinese Dialect Volume" as recording material, recorded in a professional studio with professional equipment at a 22050 Hz sampling rate. After segmentation and IPA annotation in Praat, the audio was framed with 25 ms frame length and 12.5 ms frame shift and saved. For acoustic feature extraction, MFCC (16th order), LPC (16/24/32 orders), and average intensity were concatenated into an 89-dimensional vector, with features from the previous two frames added to expand to 267 dimensions, and finally z-score standardized before input to the neural network.

4　Neural Network Structure and Praat-based Experimental Platform Model

This section introduces the 3-layer fully connected neural network structure built into Praat, with feature numbers set to 2048, 267, and 4 respectively for each layer. Through linear transformation and nonlinear activation functions processing input features, it finally outputs a 4-dimensional vector predicting the probabilities of voiceless, front high vowel, back high vowel, and other voiced sounds, with the rule that voiceless probability greater than 0.2 is judged as a voiceless segment. The model construction process based on the Praat platform is also detailed: creating "PatternList" objects storing 267-dimensional acoustic features and "Categories" objects storing 4 types of speech labels, then completing neural network construction through the "To FFNet" function. This solution reduces computational costs while ensuring recognition capability, making it suitable for experimental phonetics research on Chinese dialects.

5　Neural Network Training and Testing

This section introduces the complete process of neural network training and testing. For dataset division, Beijing and Huilai dialect points were used as the test set, and the remaining dialect points as the training set. Training parameters included: maximum training epochs of 2,400, default value for cost function decrease threshold, and cross-entropy as the cost function. Testing showed that the model achieved 89.67% accuracy in speech frame type recognition on the test set, and the recognition results were demonstrated using the Beijing Mandarin word "you" as an example, showing that the model can effectively identify voiceless segments, front high vowel segments, and back high vowel segments, with only individual deviations in transition segments.

6　Formant Extraction Based on Adaptive LPC Order

This section investigates the method of achieving precise formant extraction through adaptive adjustment of LPC order based on neural network models. Differentiated LPC order parameters were set for different speech types: 16th order for front high vowels to prevent "spurious formants," 30th order for back high vowels to avoid "formant merging," 22nd order for other voiced sounds, and no formant extraction for voiceless segments. Through comparative analysis of formant trajectory diagrams of four single-character audio recordings in Beijing Mandarin—"yi," "wu," "you," and "yao"—the limitations of traditional fixed-order LPC algorithms were verified: low orders are suitable for front high vowels but cause formant merging in back high vowels, while high orders improve back high vowel extraction but introduce spurious formants. The adaptive method proposed in this paper achieved accurate formant extraction in both monophthong and diphthong scenarios. Experimental results show that the formant trajectories extracted by this method are concentrated in distribution, have frequency values consistent with objective facts, and clearly present the dynamic change process of rhymes in diphthong syllables. Additionally, the acoustic phenomenon of concentrated trajectory points in peripheral vowel regions and instability in transition segments in diphthong syllables is explained from the perspectives of the "principle of clarity" and speech quantum theory.

7　Conclusion

This section discusses the inherent contradiction in traditional fixed-order LPC formant extraction algorithms when processing front and back high vowels: low orders easily cause "formant merging" in back high vowels, while high orders easily produce "spurious formants" in front high vowels. To address this problem, this paper designed and trained an adaptive LPC order model based on Praat's built-in neural network, which dynamically identifies speech types and sets optimal LPC orders on a frame-by-frame basis, significantly improving formant extraction accuracy and dialect research efficiency. This method requires no expensive hardware, has good popularization potential, and has broad application prospects in automatic recognition of Chinese dialects and underlying analysis of speech signals.