Glossary

Technical terms used in pstrain and acoustic model training.

Model Types

CI (Context-Independent)

Models where each phone is modeled independently of its neighboring phones. Also called monophone models. Simpler and faster to train, but less accurate than CD models. CI models are typically trained first and used to bootstrap CD training.

CD (Context-Dependent)

Models where each phone is modeled based on its surrounding context (typically the preceding and following phones). Also called triphone models. More accurate but require more training data. Example: the “AE” in “cat” is modeled differently than the “AE” in “bat”.

Tied vs Untied

Untied: Each triphone state has its own parameters. Leads to data sparsity problems since many triphones are rare.
Tied: Similar triphone states share parameters via decision tree clustering. Reduces the number of parameters while maintaining context sensitivity.

Core Concepts

HMM (Hidden Markov Model)

A statistical model where the system being modeled is assumed to be a Markov process with hidden states. In speech recognition, HMMs model the temporal evolution of speech sounds. Each phone is typically modeled as a 3-5 state HMM.

GMM (Gaussian Mixture Model)

A probability distribution modeled as a weighted sum of Gaussian distributions. Used to model the emission probabilities of HMM states. Each state can have multiple Gaussian components (mixture components).

State

A single position in an HMM. Phones are typically modeled with 3 emitting states (beginning, middle, end) plus entry and exit states.

Senone (Tied State)

A unique HMM state after state tying. Multiple triphone states that behave similarly are clustered together and share the same senone. The number of senones determines model size.

Features

MFCC (Mel-Frequency Cepstral Coefficients)

The standard acoustic features for speech recognition. Derived from the spectrum of audio, compressed using the mel scale (which approximates human hearing) and decorrelated using the discrete cosine transform.

Cepstrum

The inverse Fourier transform of the log spectrum. Separates the vocal tract response from the excitation signal.

Delta / Delta-Delta

First and second derivatives of features over time. Capture dynamic information about how features change. A 13-dimensional MFCC with deltas becomes 39-dimensional (13 + 13 + 13).

CMN (Cepstral Mean Normalization)

Subtracting the mean of cepstral features to reduce channel effects (microphone, room acoustics). Can be done per-utterance (batch) or with a running average (live).

AGC (Automatic Gain Control)

Normalizing audio amplitude to reduce volume variation effects.

Training Algorithms

Baum-Welch

An Expectation-Maximization (EM) algorithm for training HMM parameters. Iteratively re-estimates model parameters to maximize the likelihood of the training data. Also called forward-backward algorithm.

Viterbi

Algorithm for finding the most likely state sequence through an HMM given observations. Used during alignment and decoding.

Forced Alignment

Using a known transcription to determine the exact timing of phones in an audio file. The Viterbi algorithm finds the best alignment of the transcription to the audio.

Model Parameters

Means / Variances

The parameters of Gaussian distributions. Each Gaussian component has a mean vector and variance (or covariance) describing its center and spread in feature space.

Mixture Weights

The relative weights of Gaussian components in a GMM. Must sum to 1.0 for each state.

Transition Matrices (tmat)

Probabilities of transitioning between HMM states. Typically include self-loops (staying in the same state) and forward transitions.

Density

A single Gaussian component. “4 densities per state” means each state’s GMM has 4 Gaussian components.

Files and Formats

mdef (Model Definition)

Defines the structure of the acoustic model: which phones exist, how many states each has, and the mapping from triphones to tied states.

ctl (Control File)

A list of utterance IDs, one per line. Used to specify which files to process.

fileids

Same as ctl file - a list of utterance identifiers.

transcription

Text file mapping utterance IDs to their word transcripts. Pstrain accepts Sphinx [<s>] words [</s>] (utterance_id) with optional sentence markers and dict-style utterance_id words; see transcript input formats.

feat.params

The complete training-time front-end record carried by a trained model. Pstrain’s Decoder, Aligner, and package_model entry points require it, but init and training do not while they are still constructing a model.

At decode time, the trained model’s feat.params is authoritative by default over the active schema profile because it records how that model’s features were actually made. PocketSphinx reads it after pstrain’s pre-initialization assignments, so a value in the file wins. Aligner likewise uses the record’s CMN values when its cmn and cmninit arguments are omitted; passing either argument explicitly is a deliberate override. A complete file defines all 19 schema fields plus the unit_area and round_filters training-engine invariants; none is optional. Pstrain validates that inventory before its supported decode, align, and package entry points continue. A source gate rejects direct use of the pip PocketSphinx Decoder in the pstrain package, but code outside the package can bypass this contract and is not covered by the guarantee.

sendump

Precomputed senone dump file for faster model loading during decoding.

Training Stages

Flat Initialization

Creating initial model parameters before training. “Flat” means all parameters start with the same values (uniform mixture weights, global mean/variance).

Convergence

When training iterations stop improving significantly. Measured by change in log-likelihood between iterations.

Iteration

One complete pass through the training data with the Baum-Welch algorithm. Training typically runs 8-20 iterations until convergence.

Gaussian Splitting

Increasing model capacity by splitting each Gaussian into two. Start with 1 density, train to convergence, split to 2, train again, split to 4, etc.

Decision Trees

Question

A binary test about phonetic context. Examples: “Is the left phone a vowel?”, “Is the right phone a nasal?”

Quest File

File containing all the phonetic questions used for decision tree building.

Pruning

Removing branches from a decision tree to prevent overfitting. Controlled by a threshold on the minimum improvement required to keep a split.

Abbreviations Reference

Abbrev	Meaning
CI	Context-Independent
CD	Context-Dependent
HMM	Hidden Markov Model
GMM	Gaussian Mixture Model
MFCC	Mel-Frequency Cepstral Coefficients
CMN	Cepstral Mean Normalization
AGC	Automatic Gain Control
EM	Expectation-Maximization
LDA	Linear Discriminant Analysis
MLLT	Maximum Likelihood Linear Transform
MLLR	Maximum Likelihood Linear Regression
MAP	Maximum A Posteriori (adaptation)
BW	Baum-Welch (algorithm)
tmat	Transition Matrix
mdef	Model Definition
ctl	Control (file)