Phoneme Recognition

Simple implementation of DNN for phoneme recognition

Dataset

• TIMIT
• TIMIT(Spikegram)

Requirement

• python 3.6.8
• tensorflow==1.13.1
• numpy
• matplotlib
• librosa

Run

1. Spikegram

python3 run_spikegram.py

But you don't have TIMIT(spikegram).

2. MFCC

python3 run_mfcc.py

3. Spectrogram

python3 run_spectrogram.py

4. Melspectrogram

python3 run_melspectrogram.py

Result

1. Broad class

		Spikegram	MFCC	Spectrogram	Melspectrogram
Obstruent	Stops	57.38	50.97	50.45	49.18
Obstruent	Affricate	42.19	30.61	36.06	35.74
Obstruent	Fricative	70.66	66.93	66.98	67.23
Sonorant	Glides	55.14	56.59	55.98	55.43
Sonorant	Nasals	59.15	62.36	60.39	60.09
Sonorant	Vowels	53.05	53.38	52.44	53.70
Others		92.24	91.96	91.79	91.94

2. Voice & voiceless

	Spikegram	MFCC	Spectrogram	Melspectrogram
Obstruent	65.76	61.06	61.23	61.06
Sonorant	54.12	54.99	53.99	54.75
Others	92.24	91.96	91.79	91.94

3. Non-mute & mute

	Spikegram	MFCC	Spectrogram	Melspectrogram
Non mute	57.49	56.77	56.11	56.61
mute	92.24	91.96	91.79	91.94

4. Total

	Spikegram	MFCC	Spectrogram	Melspectrogram
Total	65.26	65.50	64.96	65.37

detail

Author

Han Seokhyeon