Choosing a Learning‑Rate Schedule for Kaldi nnet3 Discriminative Training
Learn how to configure a piecewise‑constant learning‑rate decay for Kaldi nnet3 discriminative training to improve convergence and WER.
Loading…
ReadMeFeed / Community knowledge
#nnet3
Real questions. Useful conversations. Find the people who know your stack.
kaldiDiagnosing and Fixing Common GPU‑Accelerated Kaldi nnet3 Training Issues
A step‑by‑step diagnostic guide for Kaldi nnet3 GPU training: identify stalls, OOM errors, high loss, crashes, and non‑deterministic results, with concrete checks, fixes, and escalation thresholds.
Kaldi nnet3 Online Decoding: Configuration Parity and Serving Reality
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 online decoding with nnet3 chain models
Use prepare_online_decoding.sh and online2-wav-nnet3-latgen-faster for low-latency streaming ASR with matching features and graph.
Getting Kaldi’s nnet3 TDNN to Train on GPU: A Practical Walk‑through
Learn how to compile Kaldi with CUDA, prepare LibriSpeech data, run a TDNN training script, and verify GPU usage while understanding memory and version trade‑offs.
Kaldi nnet3 Input Feature Normalization: Architecture and Operational Checks
An architecture note on Kaldi nnet3 input feature normalization: requirements, minimal design, trust boundaries, operational checks, and failure modes for speech pipelines.
Moving Beyond Hybrid: Implementing LF-MMI Chain Models in Kaldi
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.
Optimizing Acoustic Modeling with Kaldi's Lattice-Free MMI (Chain Models)
Learn how Kaldi's Lattice-Free MMI (Chain models) eliminates the memory bottleneck of lattice generation to provide faster, more accurate acoustic modeling.
Kaldi nnet3 Upgrade Rollback: Restoring Original Training Scripts After a Failed Build
Goal Recover a clean Kaldi environment after an unsuccessful upgrade of the nnet3 training scripts, preserving all existing data and model files. Context & Constraints The upgrade was performed from Kaldi 4.2 to 4.3. The build process terminated with a non‑fatal error, leaving the src directory partially updated. All training data and pre‑trained models