About

• SPIRE VCV is a database of speech production, which includes simultaneous acoustic and electromagnetic articulography data collected from speakers of non-native/Indian English.
  
• Stimuli comprises of non-sense symmetrical VCV (Vowel-Consonant-Vowel) utterances as part of the sentence "speak VCV today" in three different speaking rates: slow, normal, and fast with 3 repetitions each.
  
• The VCV utterances consist of the combination of 17 consonant sounds namely:
  C = { /b/, /ch/, /d/, /f/, /g/, /jh/, /k/, /l/, /m/, /n/, /ng/, /p/, /r/, /s/, /t/, /v/, /z/ }
  
• And 5 vowel sounds:
  V = { /a/, /e/, /i/, /o/, /u/ }
  
• Ten non-native English speakers: 5 female and 5 male in the age range of 18 to 27 years with no speech related disorders.
  
• Recordings were made in the sound damped studio at the SPIRE Labs speech recording facility. Acoustic and articulatory data were recorded directly to the computer and carefully synchronized.
  
• VCV boundaries were manually annotated. Read more
  

Recording and Setup

• Articulatory movements were recorded using a 3D Electromagnetic Articulograph. (EMA) AG501.
  
• A t.bone EM9600 shotgun unidirectional electret condenser microphone was placed near the subject to record the audio data synchronously with the articulatory data.
  
• Audio:
  
  • originally recorded at 48 kHz then downsampled to 16 kHz.
• Articulatory data:
  
  • Sampled at 250 Hz.
  • A 10th-order lowpass Chebyshev Type II filter with 40Hz cut-off frequency and 40 dB of stopband attenuation was used to low-pass filter the articulatory movement recording to remove the high-frequency noise resulting from EMA measurement error.
  • Sensor placement
  • 6 sensors were placed on the different speech articulators namely:
  • Upper Lip
  • Lower Lip
  • Jaw
  • Tongue Tip
  • Tongue Body
  • Tongue Dorsum
  • Sensors were also placed behind the left and right ear for the purpose of head movement correction.
  • Each of these 6 sensors captures the movements of the articulators in 3D space, resulting in eighteen articulatory features
• Instructions to speaker:
  
  • All speakers were college going students fluent with reading, writing and speaking English coming from different regions of India with different native language backgrounds.
  • Speakers were given prior training to increase speaking rate gradually during the main recording.
  • A GUI produces the stimuli to be uttered on screen and the user pronounces it for each of the three different speaking rates, namely slow, normal/habitual, and fast and three repetitions each.
• V-C-V Boundary annotation:
  
  • The VCV boundaries were manually annotated by a team of four members.
  • These boundaries were marked using an in-house built MATLAB annotation tool by observing the wideband spectrogram, the raw waveform and glottal pulses obtained using praat.
  • For unvoiced consonants: the last glottal pulse in the V1 region was considered for marking the onset of the C region, and the first glottal pulse at the start of V2 region was for considered marking the end of C-region, in tandem with the spectrogram.
  • For voiced consonants: the spectrogram with the formants and time domain waveform were considered for marking the consonant start and end boundaries.
  • For ambiguous cases, a unanimous call was then taken for the boundary marking after an internal discussion among the annotators.

Speaker Information

#	Subject	Age	Gender	Native Language
1	F1	22	Female	Bengali
2	M1	21	Male	Tulu
3	F2	27	Female	Bengali
4	M2	20	Male	Bengali
5	F3	23	Female	Tamil
6	M3	21	Male	Tamil
7	F4	20	Female	Kannada
8	M4	23	Male	Tamil
9	F5	21	Female	Malayalam
10	M5	20	Male	Hindi

Data