Train ASR and TTS models with k2, icefall, and lhotse pipelines.
npx skills add https://github.com/jayll1303/AIEKit --skill k2-training-pipeline
Please help me install this Agent Skill. Skill: k2-training-pipeline Source: https://github.com/jayll1303/AIEKit/tree/main/.kiro/skills/k2-training-pipeline Command: npx skills add https://github.com/jayll1303/AIEKit --skill k2-training-pipeline
This skill removes the complexity of building end-to-end speech training pipelines by orchestrating data preparation, model training, fine-tuning, and export for ASR and TTS systems so you can produce deployment-ready models with predictable steps and validations.
Prepare data with lhotse, run icefall training for your chosen recipe, then export the trained model to ONNX and validate it with the recipe's ONNX test script.
High-intent search queries and answers about installing and using this skill.
To train speech recognition models with k2 and icefall, you prepare data manifests using lhotse, run icefall training recipes for architectures like Zipformer or Conformer with k2 loss computation, then export the trained model to ONNX for deployment.
Yes, you can fine-tune Zipformer transducer models on custom corpora by applying the icefall training pipeline, which supports fine-tuning workflows alongside standard k2 loss functions like CTC, LF-MMI, and pruned RNN-T.
You export ASR models to ONNX by following ONNX export conventions within the icefall pipeline, separating encoder, decoder, and joiner components, then validating the exported files using the recipe's ONNX test script for sherpa-onnx compatibility.
The best way to prepare speech datasets for icefall training is using lhotse to create manifests, cuts, and extract features, ensuring your audio data meets the required manifest formats for both common corpora and custom datasets.
Yes, the k2 training pipeline supports TTS models including VITS, allowing you to handle data preparation, training, and ONNX export for both automatic speech recognition and text-to-speech projects within the same workflow.
When training Conformer models with k2, available loss functions include Connectionist Temporal Classification (CTC), Lattice-Free Maximum Mutual Information (LF-MMI), and pruned RNN-T loss computation integrated through icefall recipes.
Moshe-shipCollect and clean Arabic web content into JSONL training data.
openclaw-easyConvert input text to speech locally using sherpa-onnx and ONNX models.
Arry8Convert text to speech locally using sherpa-onnx with a runtime and model directory.
galeepLoad pre-trained Transformer models from the Hugging Face Hub for NLP, vision, audio, and multimodal tasks.
