ReadMeFeed / Community knowledge
Real questions. Useful conversations. Find the people who know your stack.
Learn how Kaldi uses WFST composition to build HCLG graphs, the role of determinization and minimization, and how to manage the memory trade-offs of large language models.
Deploy Kaldi’s chain models with online decoding: a step‑by‑step guide covering feature extraction, i‑vector adaptation, endpointing, and common pitfalls.
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.
Explore the trade‑offs between Static and Dynamic decoding in Kaldi. Learn how HCLG graphs impact ASR latency, memory usage, and grammar flexibility.
Learn how to set up Kaldi’s online decoding pipeline, run a worked example with the tedlium2 recipe, and verify latency and accuracy trade-offs.
kaldi