ReadMeFeed / Community knowledge
Real questions. Useful conversations. Find the people who know your stack.
Kaldi's chain models use LF-MMI with a single compiled denominator FST instead of per-utterance lattices or frame-level alignments. Here's what the pipeline actually looks like and how to verify it.
Learn how to run Kaldi's align utility to force‑align speech audio to a transcript, generate phone‑level timestamps, and verify the results.
Learn how Kaldi's Lattice-Free MMI (Chain models) eliminates the memory bottleneck of lattice generation to provide faster, more accurate acoustic modeling.
Kaldi's chain models train neural networks directly with lattice-free MMI at one label per 30 ms. Here's why the reduced frame rate and denominator FST design matter, and how to set up the workflow.
Learn how to reliably train Kaldi’s LF‑MMI model, from GMM‑based lattice generation to hyper‑parameter tuning, with a concrete example and practical diagnostics. Reduce WER, avoid common pitfalls, and get actionable steps for production pipelines.
The Kaldi online nnet3 decoder relies heavily on relative paths for referencing the acoustic model, final.mdl , and HCLG.fst . While these configurations function correctly when the process is executed from the local experiment or repository root, they frequently fail in production deployments where the working directory is dynamic or restricted. The toolkit
kaldi