Kaldi fbank lifter coefficient behavior across versions

Question

ContextKaldi's compute-fbank-feats added the --lifter flag to apply HTK-style liftering to filterbank coefficients. Earlier releases computed raw filterbank energies without this option, while current releases default to a lifter coefficient of 22 when the flag is present.Unresolved behaviorAcoustic models trained on features extracted with different lifter settings (or without liftering) can exhibit mismatched input distributions. The documentation does not specify whether a model trained on liftered features expects the same coefficient at inference time, or whether omitting --lifter entirely is equivalent to a coefficient of zero.ConstraintsModel checkpoints lack metadata recording the lifter coefficient used during training.Re-extracting features for a large corpus with varying --lifter values is computationally expensive.Downstream components (e.g., CMVN stats, i-vector extractors) assume a fixed feature dimension but not a fixed spectral tilt.QuestionsDoes the --lifter coefficient need to match exactly between training and decoding, or is the effect limited to a linear scaling that CMVN absorbs?When --lifter is omitted, does Kaldi apply a coefficient of zero or retain a previous default?Is there a recommended practice for recording the lifter setting alongside model artifacts to avoid silent mismatch?

Zoha · Answer

The Short AnswerThe lifter coefficient must match exactly between training and decoding. While Cepstral Mean and Variance Normalization (CMVN) handles linear shifts and scaling, it cannot compensate for the non-linear spectral tilt introduced by liftering. When the --lifter flag is omitted, Kaldi applies a coefficient of zero, resulting in raw filterbank energies without any transformation.

Technical BreakdownnIn Kaldi, liftering is a weighting process applied to filterbank coefficients to reduce correlation between adjacent bins and suppress high-frequency noise. The formula typically follows the HTK-style approach:
c_tilde(n) = c(n) / (1 + L/28 * (n - center)^2)
where L is the lifter coefficient (defaulting to 22 in modern Kaldi) and n is the coefficient index.

Distribution Mismatch: If a model is trained with --lifter 22 and decoded without it, the high-frequency coefficients will have significantly higher variance than the model expects. This creates a distribution shift that the acoustic model cannot resolve, leading to a degraded Word Error Rate (WER).
  CMVN Limitations: CMVN normalizes the mean and variance of each dimension. Because liftering applies a frequency-dependent weighting, a simple linear normalization cannot "undo" the spectral tilt change.

Verification and Best Practices
To verify the behavior in your specific Kaldi build and ensure consistency, follow these steps:

Confirm Zero-Coefficient: Run compute-fbank-feats once with --lifter 0 and once omitting the --lifter flag entirely. Compare the output .scp files; they should be numerically identical.
  Test Sensitivity: If you are uncertain of a model's origin, decode a small test set both with and without the lifter flag. A sharp drop in acoustic confidence scores usually indicates the model was trained with the lifter active.
  Metadata Tracking: Since Kaldi does not store the lifter coefficient in the model tree, the standard practice is to include the feature extraction parameters in your recipe's yaml or a custom metadata file alongside the checkpoint directory.

Kaldi fbank lifter coefficient behavior across versions

Context

Unresolved behavior

Constraints

Questions

1 answer

The Short Answer

Technical Breakdown

Verification and Best Practices

Your perspective belongs here.

0 question comments