Choosing Between Hybrid HMM-DNN and End-to-End Architectures in Kaldi
A technical guide on choosing between Hybrid HMM-DNN and End-to-End architectures in Kaldi, focusing on data requirements, vocabulary flexibility, and WER validation.
Loading…
ReadMeFeed / Community knowledge
Real questions. Useful conversations. Find the people who know your stack.
A technical guide on choosing between Hybrid HMM-DNN and End-to-End architectures in Kaldi, focusing on data requirements, vocabulary flexibility, and WER validation.
Learn how to balance the G (Grammar) component of Kaldi's HCLG graph to optimize the trade‑off between Word Error Rate (WER) and memory consumption during decoding.
Batch versus online CMVN in Kaldi recipes Kaldi recipes do not document a single universal default for cepstral mean and variance normalization. In batch pipelines, compute-cmvn-stats typically produces a transform matrix that apply-cmvn later reuses for development and test data. Some recipes compute global statistics from the training set, while others der
A Kaldi chain/TDNN acoustic model decodes correctly inside the original recipe directory, but the same model fails when deployed in a separate production pipeline. The nnet3 decoding binaries abort with low-level errors around the i-vector input — missing final.ie or a feature-dimension complaint from the nnet3 computation — rather than any explicit configur
Goal Determine a cost‑effective deployment strategy for Kaldi on low‑traffic speech workloads by balancing the real‑time online2 decoding pipeline with aggressive nnet3 model pruning. Constraints & Uncertainty The online2‑wav‑nnet3‑latency script defaults to a 30 ms stride, yet the trade‑off between this stride setting, decoding latency, and word error r
In Kaldi's FST-based decoding architecture, the composition of H, C, L, and G transducers results in a search graph where weight pushing is applied to optimize pruning during the Viterbi beam search. This process is intended to move weights as far forward as possible toward the start of the graph. There is a specific design consideration regarding how this o
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 maintains a distinct boundary between legacy HMM-based models and modern end-to-end neural architectures. Each approach requires a specific graph-processing pipeline to ensure that the feature extractor and the decoder are aligned. When transitioning a deployment from HMM-based decoding to neural architectures, the configuration of the scf file and the