GigaAM

Why we included this project

GigaAM is a Conformer-based acoustic model family with state-of-the-art results on Russian speech recognition, and a newer multilingual line extends the same approach to 70+ languages. The repo ships pretrained CTC and RNN-T checkpoints you can load and call transcribe() on for short clips, or use long-form transcription with word-level timestamps for longer recordings. The same encoder also backs an emotion recognition model, so one audio pipeline can handle both transcription and affective analysis. Fine-tuning runs through PyTorch Lightning, and ONNX export plus Triton Inference Server support cover the production serving side. If you're building call-center analytics, voice assistants, or multilingual transcription, this is a practical MIT-licensed starting point.