Decoder-Error-Based Similarity Matrices

Samples for decoder-error based similarity matrices for Spanish, English and Basque are below.

The scores in our similarity matrix represent error-percentages from our phone-decoder, normalized into a 1-1000 range. A score of -500 i.e.,

1/2 × ( 0 – max( {ScoreRange} ) was stipulated for cases where no confusion had taken place for a phoneme-pair.

For instance, a score of 85 for phoneme pair [θ, f] in Spanish is computed from the observation that 8.5% of the occurences of /θ/ were decoded as /f/.

The error counts were obtained with HTK's HResults, when aligning the phonetic recognition and the G2P transcription for sequences of ca. 25000 phonemes in Spanish and ca. 12700 phonemes in English.

As tor the decoders, depending on the study we've used a monophone or a triphone decoder. The differences in the scoring matrices were not large when using either decoder.

Monophone decoder

It was trained with HTK. The acoustic models had three left-to-right emitting states using 32 Gaussian mixture components. The parametrization of the signal consisted of 18 Mel-Frequency Cepstral Coefficients plus the enery and their delta and delta-delta coefficients, using 16-bit PCM audios sampled at 16 KHz.

Triphone decoder

Cross-word triphone models were trained with HTK. The parametrization of the signal consisted of 18 Mel-Frequency Cepstral Coefficients plus the energy and their delta and delta-delta coefficients, using 16-bit PCM audios sampled at 16 KHz.