ReadMeFeed / Community knowledge
Real questions. Useful conversations. Find the people who know your stack.
Kaldi's nnet3 online decoding enables streaming ASR, but silent accuracy loss from configuration drift between training and serving is the real production risk. This post covers feature parity checks, chunk-size vs. lookahead trade-offs, and the serving wrapper you'll actually need to build.
Kaldi's Chain/LF-MMI models replaced the hybrid DNN/HMM stack with a single TDNN trained end-to-end. This guide covers the architectural shift, training objective changes, data-prep differences, and practical migration steps with a LibriSpeech worked example.
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.
Learn how to implement LF-MMI chain models in Kaldi to eliminate frame-level alignment bottlenecks and improve ASR accuracy using TDNNs and i-vectors.
Learn how Kaldi’s chain models with TDNN‑F and lattice‑free MMI replace legacy DNN/HMM hybrids, cutting training time and model size while keeping or improving word error rates.
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 Kaldi’s i‑vector extractor turns raw audio into speaker‑adapted features, how to add it to your recipe, and the trade‑offs you’ll face in real‑time decoding.
Learn how fMLLR reduces speaker variability in Kaldi, see a minimal script to load and apply the transform, and verify its correctness before deploying in production.
When deploying Kaldi for low-traffic workloads on resource-constrained hardware, memory consumption is primarily dictated by the size of the HCLG.fst graph loaded into RAM. To minimize infrastructure costs, developers often utilize on-the-fly rescoring to reduce the initial graph footprint. However, reducing hardware resources necessitates a careful balance
kaldi