CMU Sphinx Open Source Models

Table of Contents

1. Language Model Resources
2. Language models for HUB4 (broadcast news)
3. General Notes on Acoustic Models
1. PocketSphinx Specific Notes
2. Sphinx-3 Specific notes
4. Acoustic Models for Sphinx-3 and PocketSphinx
5. Models for Sphinx-2 (obsolete)

Language Model Resources

• Keith Vertanen's English Gigaword Language Models are suitable for general purpose dictation.
• For small dialog and command-and-control tasks, you can use the Sphinx Knowledge Base Tool.

Language models for HUB4 (broadcast news)

• Dictionary
• Filler dictionary
• Phone list
• Language model in DMP and gzip format
• Vocabulary list used for creating the language model

General Notes on Acoustic Models

All models use the new "40 phone" phoneset and are compatible with CMUDict 0.6d and 0.7, as well as the Sphinx Knowledge Base Tool. These models are released under the same permissive license as Sphinx-3. Although the databases they are trained from are not freely redistributable, as far as we know, this does not affect models trained on them.

These models are packaged for use with Sphinx-3 and PocketSphinx. Unpack them using tar and take note of the directory that is created by doing so. To use them, you will specify this directory in your configuration file, or on the command line, using the -hmm argument. This will ensure that the proper feature extraction parameters are picked up by the decoder.

PocketSphinx Specific Notes

The mixture_weights file is not necessary in order to use the semi-continuous models with PocketSphinx. Since it is quite large, you may choose to omit it in cases where storage is limited. We have included it because it is (currently) necessary in order to do acoustic model adaptation.

If you use these continuous density models with PocketSphinx you may encounter segfaults. This is a known bug which will be fixed in the next release. To solve it, you need to edit the noisedict file and remove all lines where the pronunciation of a word contains phones other than SIL or noise phones (ones starting and ending with +). Alternately, here is a corrected noise dictionary. Copy it over the noisedict file.

Sphinx-3 Specific Notes

While Sphinx-3 supports MLLT feature transformations, version 0.7 does not find them automatically in the acoustic model directory. Therefore, for this version, in order to use the wideband WSJ models below, you will need to add this argument to the configuration or command line:

-lda wsj_all_cd30.mllt_cd_cont_4000/feature_transform

If you aren't running the directory from the same directory as the model, you will need to add the full path to it.

Note also that if you intend to use the semi-continuous models with Sphinx-3, you must add these arguments to your configuration:

-feat s2_4x -senmgau .s2semi.

Acoustic Models for Sphinx-3 and PocketSphinx

• Communicator (dialog system) acoustic models - for narrowband (8kHz) telephone speech
• 6000 senone semi-continuous models (for PocketSphinx)
• 4000 senone, 64 Gaussian continuous density models (for Sphinx-3)

Important note: These continuous-density acoustic models are very large and will not run in real-time with the standard set of parameters. Unfortunately, Sphinx-3 has a large number of tunable options for speeding things up, and tuning them is something of a black art. We suggest starting with the following settings:

-subvq      Communicator_40.cd_cont_4000/subvq
-beam       1e-60
-wbeam      1e-40
-ci_pbeam   1e-8
-subvqbeam  1e-2
-maxhmmpf   2000
-maxcdsenpf 1000
-maxwpf     8
-ds         2

• WSJ1 (dictation) acoustic models - for wideband (16kHz) microphone speech
• 5000 senone semi-continuous models (for PocketSphinx)
• 4000 senone, 32 Gaussian continuous density models with MLLT transform

Important note: See note above about the -lda argument to Sphinx-3
Important note #2: These continuous-density acoustic models are very large and will not run in real-time with the standard set of parameters. Unfortunately, Sphinx-3 has a large number of tunable options for speeding things up, and tuning them is something of a black art. We suggest starting with the following settings:

-subvq      wsj_all_cd30.mllt_cd_cont_4000/subvq
-beam       1e-80
-wbeam      1e-60
-subvqbeam  1e-2
-maxhmmpf   2500
-maxcdsenpf 1500
-maxwpf     20

• WSJ1 (dictation) acoustic models - for narrowband (8kHz) microphone speech
• 5000 senone semi-continuous models (for PocketSphinx)
• 3000 senone, 32 Gaussian continuous density models (for Sphinx-3)
• HUB4 (broadcast news) acoustic models - for wideband (16kHz) speech
• Information about models (out of date)
• 4000 senone, 8 Gaussian continuous density models
• 6000 senone, 8 Gaussian continuous density models

Models for Sphinx-2 (obsolete)

Important note: These models use an older (44-phone) phoneset, so you must use the Advanced Sphinx Knowledge Base Tool instead (select "Reduced (Sphinx_44)" in the dictionary and language model parameters).

• Acoustic models
• Language models