Phones dataset for speech recognition (not telephone number)

Question

In phonetics a phone is a unit of speech sound.

I am looking for a dataset of phone waveforms for speech recognition. There are about 17 vowels and 17 consonants for English phones. I tried looking everywhere, but couldn't find a reliable source.

Any help would be appreciated.

Answer 1

1. You can find US and British english language phones at university linguistic departments such as Berkeley, Ca. Though, what is useful for speech recognition are the features in statistical acoustic models that are extracted from a large amount of annotated audio recordings. These acoustic model features involve hidden markov models and Mel-Frequency Cepstral coefficients. (MFCC tutorial by James Lyons below) Interpreting and understanding these feature files is complex. The CMU Sphinx speech recognition project provides different acoustic models such as AN4 and RM1. The distribution of some CMU audio databases such as RM1 are restricted. The AN4 database includes the audio recordings and utilizes 34 phones.

Berkeley Linguistics Phonetics
(click on phone, then left click for audio, right click for spectrogram)
https://corpus.linguistics.berkeley.edu/acip/

CMU Sphinx acoustic models
http://www.speech.cs.cmu.edu/databases/an4/index.html

CMU Sphinx dictionary
https://github.com/cmusphinx/cmudict
cmudict.dict
yes Y EH1 S
no N OW1

cmudict.phones (sample of 39 phones vs 34 phones for AN4)
Y semivowel
EH vowel
S fricative
N nasal
OW vowel

cmudict.symbols
Y
EH1
OW1

2. Voxforge provides open source corpus, audio recordings and acoustic models for multiple speech recognition engines in multiple languages. CMU Sphinx, ISIP, Julius, and HTK http://www.voxforge.org/home
   https://github.com/julius-speech/julius
   https://www.isip.piconepress.com/projects/speech/
   http://htk.eng.cam.ac.uk/

3. James Lyons provides an excellent Mel Frequency Cepstral Coefficient (MFCC) tutorial along with python code.
   http://www.practicalcryptography.com/miscellaneous/machine-learning/guide-mel-frequency-cepstral-coefficients-mfccs/
   https://github.com/jameslyons/python_speech_features

4. Adam Coates of Baidu Research and Stanford gave an excellent presentation about improving the acoustic model.
   http://cs.stanford.edu/~acoates/ba_dls_speech2016.pdf
   https://github.com/baidu-research/ba-dls-deepspeech
   http://www.openslr.org/12/ (corpus, audio recordings, acoustic models)