ReadMeFeed / Community knowledge
Real questions. Useful conversations. Find the people who know your stack.
Kaldi decoders emit word lattices, making two-pass decoding practical: a fast first pass prunes the search space, and a heavier second pass rescores survivors. Learn how to diagnose recall loss with oracle WER.
Learn how Kaldi's feature splicing expands temporal context for acoustic models, the trade-offs between dimensionality and compute, and how to verify your feature dimensions.
Decide between Kaldi’s classic HMM‑GMM and modern DNN‑HMM acoustic models. Compare speed, accuracy, memory and see a step‑by‑step implementation to help you choose the right approach.
Kaldi’s online2 framework lets you run neural acoustic models in real time. This guide walks through a minimal script, explains beam‑width tuning, and shows how to balance latency, CPU usage, and accuracy on both desktops and embedded devices.
A practical architecture note on Kaldi's HMM-GMM pipeline: requirements, minimal design, data boundaries, WER checks, and failure modes.
An architecture note on Kaldi nnet3 input feature normalization: requirements, minimal design, trust boundaries, operational checks, and failure modes for speech pipelines.
Configure Kaldi speech decoding by matching MFCC features to your acoustic model, preparing the data directory, and running standard decode scripts for GMM or chain models.
A decision guide for new Kaldi projects: when a GMM-HMM baseline is enough, when to go straight to a chain TDNN recipe, and how to validate the choice with a dev-set WER comparison.
An architecture note for a minimal Kaldi online decoder: a single-threaded loop around an HCLG graph and DecodableInterface, with explicit trust boundaries, token-count and RTF checks, and the failure modes that force a redesign.
kaldi