Originally published at: Integrating NVIDIA Triton Inference Server with Kaldi ASR | NVIDIA Technical Blog
Speech processing is compute-intensive and requires a powerful and flexible platform to power modern conversational AI applications. It seemed natural to combine the de facto standard platform for automatic speech recognition (ASR), the Kaldi Speech Recognition Toolkit, with the power and flexibility of NVIDIA GPUs. Kaldi adopted GPU acceleration for training workloads early on. NVIDIA…
The Jupyter Notebook referred to in this post is not accessible. Would it be possible to include it with the github repo vs. your internal gitlab?
Upon following the instructions in the quick start guide we were able to launch the Triton server successfully, but upon running scripts/docker/launch_client.sh the client hung after outputting “Opening GRPC contextes…” (without the “done”) and before outputting “Streaming utterances…”
A quick glance at the client code appears it is likely hanging on line 273 of kaldi-asr-client/kaldi_asr_parallel_client.cc in the TritonASRClient constructor call.
Has anyone seen this issue before?
This problem appears to be on our end, result of a k8s networking issue in our on-prem k8s cluster. Was eventually able to successfully run the Triton Kaldi ASR client against the server.