ONNX conversions of ASR models with attentions enabled for output. Especially useful for word-level timestamp extraction.