salute-developers/GigaAM

Foundational Model for Speech Recognition Tasks

What it solves

GigaAM provides a family of open-source acoustic models designed for high-performance speech processing. It addresses the need for state-of-the-art (SoTA) automatic speech recognition (ASR) and emotion recognition, particularly for the Russian language and several underrepresented languages (such as Kazakh, Kyrgyz, and Uzbek).

How it works

It uses a Conformer-based foundational model (ranging from 220M to 600M parameters) pre-trained on massive amounts of speech data—up to 2 million hours for the multilingual version. The system employs different decoding strategies, including Connectionist Temporal Classification (CTC) and Recurrent Neural Network Transducer (RNNT), to convert audio into text. It also supports self-supervised learning (SSL) backbones for audio embedding extraction and specialized models for emotion detection.

Who it’s for

Developers and researchers working on speech-to-text applications, emotion analysis, and multilingual audio processing, especially those targeting the Russian-speaking region or underrepresented languages.

Highlights

  • Multilingual Support: Pre-trained on 70+ languages with high performance on Russian, Kazakh, Kyrgyz, and Uzbek.
  • End-to-End ASR: Includes v3_e2e models that support punctuation and text normalization, outperforming Whisper-large-v3 in side-by-side comparisons.
  • Deployment Flexibility: Supports ONNX export for faster GPU inference and integration with Triton Inference Server and TensorRT.
  • Versatile Tasks: Capable of audio embedding extraction, word-level timestamps, and long-form transcription via external Voice Activity Detection (VAD).
  • Fine-tunable: Models can be fine-tuned on custom datasets using PyTorch Lightning.

Related

  • Project
    snakers4/silero-models

    A collection of fast, end-to-end text-to-speech models supporting a wide variety of languages, designed for efficient execution on CPU and GPU.

  • Project
    modelscope/FunASR

    An industrial speech recognition toolkit for offline, streaming, and edge deployment, providing high-performance ASR, VAD, and speaker diarization pipelines.

  • Dispatch
    Hugging Face Open ASR Leaderboard Trends and Insights

    Hugging Face has updated the Open ASR Leaderboard with new multilingual and long-form tracks, revealing that Conformer encoders paired with LLM decoders currently lead in English accuracy while CTC/TDT decoders offer superior throughput.

  • Project
    QwenAudio/Fun-ASR

    A family of high-precision end-to-end speech recognition models optimized for noisy environments, diverse Chinese dialects, and high-throughput production deployment.

  • Dispatch
    Fine-Tune MMS Adapter Models for Low-Resource ASR

    Hugging Face announced that fine‑tuning adapter layers on Meta AI's Massive Multilingual Speech (MMS) models yields state‑of‑the‑art word error rates for low‑resource languages after only minutes of training, while being more memory‑efficient than full model fine‑tuning.