GigaAM — speech and emotion recognition

★ 831

GigaAM is a family of open-source Python acoustic models for speech and emotion recognition from the SaluteSpeech team, MIT-licensed. Reach for it when you need to transcribe call-center recordings, voice messages and meeting audio; when accuracy on Russian matters; when you need Kazakh, Kyrgyz or Uzbek recognition; or when audio must not go to a cloud service. Inside: a Conformer encoder of 220M to 600M parameters and four lines, v1, v2, v3 (pretrained on Russian speech) and multilingual (70+ languages, 2 million hours; English quality is moderate). The v3-e2e models output punctuated, normalized text; there are CTC and RNNT decoders, the GigaAM-Emo emotion model, word-level timestamps, ONNX export, Triton Inference Server support and fine-tuning. Python 3.10+ and ffmpeg are required, and weights are hosted on Hugging Face. The transcribe method handles audio up to 25 seconds; longer recordings need pyannote.audio and a Hugging Face token. The direction is speech to text, not speech synthesis. According to the authors, the e2e models win 70:30 against Whisper-large-v3 in an LLM-judged side-by-side comparison. Suits local deployment; not for voice generation.