Athena is an open-source speech processing engine implemented in C++ and built on TensorFlow 2.0 and above. The project provides end-to-end implementations for multiple speech processing tasks including automatic speech recognition, text-to-speech synthesis, voice activity detection, and keyword spotting. The repository emphasizes making speech processing accessible to both industrial applications and academic research by releasing example implementations and recipes on open-source datasets.
The core architecture supports hybrid attention and CTC-based methods for ASR with both end-to-end and streaming capabilities. For text-to-speech, Athena implements FastSpeech, FastSpeech2, and Transformer-based models. The framework includes voice activity detection functionality and keyword spotting with end-to-end and streaming variants. A notable feature is ASR unsupervised pre-training using masked predictive coding. The system supports multi-GPU training across single and multiple machines using Horovod, and includes WFST creation and WFST-based decoding implemented in C++. Deployment is supported through TensorFlow C++ for local server implementations.
The repository provides a Kaldi-free pythonic feature extractor called Athena_transform to simplify usage. Recent additions include Conformer-CTC models, Transformer-U2 architectures, CTC alignment functions, and speaker embedding capabilities. The framework has been actively developed with significant updates in 2021 and 2022, including performance optimizations for E2E ASR models, Horovod parameter adjustments for faster training, and the addition of SpecAugment functionality.
Performance results are documented across multiple datasets. For ASR, the framework achieves competitive results on AISHELL-1, HKUST, LibriSpeech, GigaSpeech, and MISP datasets using various model architectures. TTS performance is demonstrated on LJSpeech and data_baker datasets with multiple acoustic models and vocoders. VAD achieves frame error rates of 8.49 percent with DNN and 2.50 percent with MarbleNet on Google Speech Commands Dataset V2. KWS results on MISP2021 task1 show performance metrics across streaming and end-to-end configurations.
The deployment evaluation shows real-time factor measurements on CPU hardware, with CTC Prefix BeamSearch achieving RTF of 0.0283 on 10 logic cores with beam size 1. The repository maintains two major versions with Athena v2.0 as the current master branch. An associated model zoo repository provides pre-trained models for quick experimentation.
According to GitGenius activity tracking, the repository shows extended response latencies with a median of 13059.3 hours and mean of 29020.1 hours across tracked issues and pull requests. The most active labels tracked are bug and stale classifications. The project connects to related repositories including funaudiollm/cosyvoice, pytorch/pytorch, and stability-ai/stable-audio-tools through overlapping contributors. The framework is classified across multiple domains including AI frameworks, deep learning, neural networks, machine learning, research tools, and computational tools, reflecting its broad applicability in speech processing research and deployment.
