Tuva Speech Database

Tuva Speech Database was created by Max Manus AS for testing and evaluation of the speech recognition solution "Tuva" for Norwegian. The corpus consists of approximately 24 hours of recorded speech from 40 speakers of Norwegian, 36 of which speak a dialect close to the Bokmål written standard, while four speak a dialect that is closer to the Nynorsk written standard. About 70% of the material is manuscript read. The remaining 30% is spontaneous speech. The manuscripts in the manuscript read part of the corpus is for the most part composed of short news articles. 25% of the manuscripts are read by all speakers, while the remaining 75% are unique to each speaker. All punctuation (dots, commas, paragraphs etc.) are read by the speakers, and all sound recordings are orthographically transcribed in two different formats. For Nynorsk, only manuscript read speech is available. The speakers have been selected to represent a cross section of the Norwegian working population, balanced for age, gender and dialect. All recordings are made with a 48 kHz sampling frequency and 32 bit resolution with one microphone in one channel (mono). The recordings were conducted in a recording studio in Oslo.