CMU Sphinx Acoustic Models - US English

Home > Software > Sphinx WSJ training recipe > US acoustic models

Here you can download some of the US-English acoustic models I trained using my Sphinx Wall Street Journal Training Recipe.

Models are available using different amounts of training data, number of senones, continuous vs semi-continuous, HMM topologies, and number of Gaussians per state. They all are using the 40 phone set from the CMU dictionary (without stress). Except where indicated, I trained using 1s_12c_12d_3p_12dd acoustic features. All models were trained on 16kHz audio, except for the narrowband ones which were downsampled 8kHz audio.

The garbage phone models have an extra X phone added for use in garbage modelling. I replaced 10% of the words in the training transcripts with a garbage word with a pronunciation consiting of the X phone repeated for each phone in the real word's pronunciation. I'm not sure if this is the best way to do this, but it seems to work.

Training data Type Topology Senones Gaussians Size
WSJ SI-84 cont 3 states, no skips 8000 32 73MB Download
WSJ SI-284 cont 3 states, no skips 8000 32 74MB Download
WSJ all cont 3 states, no skips 4000 32 38MB Download
WSJ all cont 3 states, no skips 6000 32 56MB Download
WSJ all cont 3 states, no skips 8000 32 73MB Download
WSJ all cont 3 states, no skips 10000 32 73MB Download
WSJ all cont 5 states, skips 8000 32 74MB Download
WSJ all cont 3 states, no skips 8000 1 4MB Download
WSJ all cont 3 states, no skips 8000 2 6MB Download
WSJ all cont 3 states, no skips 8000 4 11MB Download
WSJ all cont 3 states, no skips 8000 8 20MB Download
WSJ all cont 3 states, no skips 8000 16 38MB Download
WSJ all cont 3 states, no skips 8000 64 146MB Download
WSJ all semi 3 states, no skips 8000 256 30MB Download
WSJ all semi 5 states, skips 8000 256 31MB Download
WSJ all semi, s2_4x 5 states, skips 8000 256 31MB Download
WSJ all semi, s2_4x 5 states, skips 8000 128 16MB Download
WSJ all semi, s2_4x 5 states, skips 8000 512 57MB Download
WSJ all semi, s2_4x 5 states, skips 8000 1024 106MB Download
WSJ all semi, narrowband 5 states, skips 8000 256 31MB Download
WSJ all semi, narrowband 3 states, no skips 8000 256 31MB Download
WSJ all cont, garbage phone 3 states, no skips 8000 4 11MB Download
WSJ all cont, garbage phone 3 states, no skips 8000 8 20MB Download
WSJ all cont, garbage phone 3 states, no skips 8000 16 38MB Download