GigaAM-v3

GigaAM-v3

Turns Russian speech into text, with variants that add punctuation and normalized text, built on a Conformer of about 230M parameters. For developers and teams transcribing Russian audio such as calls.

Made by ai-sage on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: GigaAM-v3 is a speech recognition model family from ai-sage, built on a Conformer encoder of 220 to 240M parameters. The encoder was pretrained on 700,000 hours of Russian speech. Five variants are offered, from the raw encoder to models that return punctuated, normalized text.

Who it suits: Developers and teams who transcribe Russian audio, including phone calls, background-music clips and everyday speech. The record also lists English. The card says the third generation does about 30 percent better on its new domains than earlier ones.

Title: Who uses it, and how
Industry | Who uses it | How they use it | Example
Customer service | Call-center teams | Transcribe recorded calls into text | A team converts a saved support call into a written record
Media | Caption editors | Draft subtitles for Russian audio | An editor transcribes a clip with background music
Accessibility | App developers | Transcribe speech with atypical characteristics | A developer adds transcription for disordered speech
Research | Speech researchers | Reuse the pretrained encoder or compare decoders | Compare the CTC and RNN-T variants on the same clips

Skip it if your audio is in a language other than Russian or English.

Where to use it:

• Python with the transformers library
• Scripts and services that transcribe audio files
• Windows, macOS or Linux for downloading the files

Use case | Who | What they get back
Call transcripts | Support teams | Text of what was said on a recorded call
Punctuated captions | Editors | Text with punctuation and normalized wording from the end-to-end variants
Research baselines | Researchers | An encoder and decoders to compare against other systems

How to set it up: You need Python and pip, plus the library versions the card recommends.

• Python and pip
• torch and torchaudio version 2.8.0
• transformers version 4.57.1
• pyannote-audio 4.0.0 and torchcodec 0.7.0
• hydra-core, omegaconf and sentencepiece, in any version

1. Install the Hugging Face hub library.
2. Download the repository.
3. Install the recommended libraries and run the Python example.

pip install -U huggingface_hub
hf download ai-sage/GigaAM-v3

How to use it: Pick a variant through the revision setting, load the model, and call transcribe on an audio file. The end-to-end variants return punctuated text.

from transformers import AutoModel

revision = "e2e_rnnt"  # can be any v3 model: ssl, ctc, rnnt, e2e_ctc, e2e_rnnt
model = AutoModel.from_pretrained(
    "ai-sage/GigaAM-v3",
    revision=revision,
    trust_remote_code=True,
)

transcription = model.transcribe("example.wav")
print(transcription)

Tool | What it does
ssl | The self-supervised encoder, pretrained on Russian speech
ctc | Speech recognition with a CTC decoder
rnnt | Speech recognition with an RNN-T decoder
e2e_ctc | CTC recognition that adds punctuation and normalizes text
e2e_rnnt | RNN-T recognition that adds punctuation and normalizes text

Example: You set the revision to e2e_rnnt and call transcribe on a recorded Russian call, example.wav. The script prints the spoken words as ready-to-read text with punctuation.

For newcomers and experts: If you are new, run the short Python example on one audio file with the e2e_rnnt variant and read what it prints. Try the other variants afterward.

For people in the field, the encoder uses a HuBERT-CTC pretraining objective, and the ASR variants add either a CTC or an RNN-T decoder. In the card's tests, the average word error rate is 8.4 for RNN-T and 9.2 for CTC, against 25.1 for Whisper. Judged by an LLM, the end-to-end models beat Whisper-large-v3 by about 70 to 30.

Audio file in
Encoder reads the speech
CTC or RNN-T decoder writes words
End-to-end models add punctuation

Good to know:

Title: Average word error rate, lower is better
GigaAM-v3 RNN-T: 8.4
GigaAM-v3 CTC: 9.2
T-One with LM: 19.4
Whisper: 25.1

• License: MIT
• Cost to run: none named, you use your own computer
• Data it reaches: it downloads files from Hugging Face, then reads the audio files you give it, and the example allows the repository's own code to run
• Not for: audio in other languages, since pretraining was on Russian speech
• Disordered speech is the hardest set in the card, at 19.2 for RNN-T

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Downloads in the last 30 days: 257,541
• Likes: 202
• License: mit
• Languages: ru, en
• Quantized versions: 44
• Fine-tunes: 15
• Tags: pytorch, gigaam, automatic-speech-recognition

Links

• Model: https://huggingface.co/ai-sage/GigaAM-v3
• Files: https://huggingface.co/ai-sage/GigaAM-v3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ai-sage/GigaAM-v3

Then follow the usage in its model card. Nothing in this kit runs by itself.