viks66/SPIRE_EMA_CORPUS

This corpus contains paired data of speech, articulatory movements and phonemes. There are 38 speakers in the corpus, each with 460 utterances.

The raw audio files are in audios.zip. The ema data and preprocessed data is stored in processed.zip. The processed data can be loaded with pytorch and has the following keys -

• ema_raw : The raw ema data
• ema_clipped : The ema data after trimming using being-end time stamps
• ema_trimmed_and_normalised_with_6_articulators: The ema data after trimming using being-end time stamps, followed by articulatory specifc standardisation
• mfcc: 13-dim MFCC computed on trimmed audio
• phonemes: The phonemes uttered for the audio
• durations: Duration values for each phoneme
• begin_end: Begin end time stamps to trim the audio / raw ema

To use this data for tasks such as acoustic to articulatory inversion (AAI), you can use ema_trimmed_and_normalised_with_6_articulators and mfcc as the data.

---

If you have used this dataset in your work, use the following refrence to cite the dataset -

Bandekar, J., Udupa, S., Ghosh, P.K. (2024) Articulatory synthesis using representations learnt through phonetic label-aware contrastive loss. Proc. Interspeech 2024, 427-431, doi: 10.21437/Interspeech.2024-1756