Note: This file must be read by anyone who intends to use the models in this package for recognition. The specifications provided here must be exactly matched in the user's setup to prevent recognition failures.
The models have been trained using 140 hours of 1996 and 1997 hub4 training data, available from the Language Data Consortium. The phoneset for which models have been provided is that of the CMU dictionary version 0.6d. The dictionary has been used without stress markers, resulting in 40 phones, including the silence phone, SIL. Adding stress markers degrades performance by about 5% relative.
The models have been trained with Mel-frequency cepstra (MFC) vectors derived from the hub4 data. Each vector is composed of 13 cepstral coefficients, 13 delta cepstra and 13 double delta cepstra. Each vector is thus 39-dimensional. The correct SPHINX name for the vectors used is "1s_c_d_dd", and this must be specified to the decoder(s) for correct usage of the acoustic models provided. The specifications for the feature set are as follows:
The models are 3-state within-word and cross-word triphone HMMs with no skips permitted between states. There is one set of models in this package, comprised of 6000 senones. Other sets are or will soon be available from the CMU Open Source Sphinx web page. A set of quantized models have also been provided with the set of models. The models have 8 Gaussians per state, and the quantized models use 4096 codewords.
The quantized models are for use with s3.2/3.3 (fast decoder), which also requires the corresponding un-quantized models during runtime (ie, both must be provided). The quantized models are labeled as .quant and are placed in the same subdirectory as the corresponding full model.
The un-quantized models can be used with the s3 continuous decoder (slow decoder).
The language model in this distribution is provided as an example of use and test for the installation, and is not meant to be used with any serious task. It is a flat unigram language model for CMU's Census database. The dictionary, also provided, is restricted to the letters of the alphabet and some additional control words.
Another language model, provided through the Sphinx web page referred above, , a simple trigram model, and which has been built for tasks similar to broadcast news, is or will soon be available from the Sphinx web page refered above. You are strongly encouraged to get this language models if you intend serious applications. The text used to build this model was taken from a variety of permitted sources, including broadcast news. The vocabulary covers 64000 words, and is listed in the file called language_model.vocabulary. The file language_model.arpaformat.gz can be used with the Sphinx-2 decoder, while the file language_model.arpaformat.DMP.Z must be used with the Sphinx-3 decoders. Note that the system will only recognize words which are within the vocabulary. See a description of the ARPA format.
Maintained by Evandro B. Gouvêa Last modified: Mon Nov 25 18:24:14 EST 2002