GigaAM-Multilingual

Name: GigaAM-Multilingual
Author: ai-sage

by ai-sage

Conformer ASR pretrained on 2M hours, strongest open-source Russian, Kazakh, Kyrgyz and Uzbek

87likes

220M / 600Mparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

ai-sage/GigaAM-Multilingual

Feature URI

Deploy GigaAM-Multilingual

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

Read the customer-hosted guide Talk to us about single-tenant

Overview

GigaAM Multilingual is a family of Conformer speech models released in July 2026, pretrained with a HuBERT-style objective on 2 million hours of speech across more than 70 languages, then fine-tuned for recognition with character-wise CTC decoders on 50,000 hours. The family ships at two sizes, 220M and 600M parameters, each as a self-supervised encoder and as a CTC recognizer.

Its own card is unusually direct about where it wins and loses, claiming the strongest open-source quality available on Russian, Kazakh, Kyrgyz and Uzbek and moderate quality on English. The published WER table backs both halves, and the English rows show Whisper large v3 ahead. Pick this model for the languages it was built for, and keep Whisper for English.

In a retrieval pipeline ASR is the step that turns audio into text, and the text is then embedded by a text model. Transcription quality sets the ceiling on every search over spoken content, which is why the choice of ASR model per language matters more than the embedding model that follows it.

Architecture

Conformer encoder pretrained with a HuBERT-style objective on 2M hours across 70+ languages, fine-tuned on 50K hours with a character-wise CTC decoder. Four published variants: ssl (220M self-supervised encoder), ctc (220M recognizer), large_ssl (600M encoder), large_ctc (600M recognizer). The repo ships custom code, so loading it requires trust_remote_code.

Mixpeek SDK Integration

// ASR runs on your side (or on a single-tenant deployment with an uploaded
// model) and the transcript is what enters Mixpeek. Embed the text with a text
// model; the audio itself stays in your bucket.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection_id: "col_your_collection",
      documents: [
        {
          document_id: "call-00412-seg-07",
          vectors: { "text-embedding": yourTranscriptVector },
          payload: {
            source_key: "calls/2026/00412.wav",
            language: "kk",
            start_ms: 184000,
            transcript: "...",
          },
        },
      ],
    }),
  },
);

Capabilities

Character-wise CTC decoding, which is greedy and fast with no language-model pass
Self-supervised encoder variants usable as a speech feature extractor rather than a transcriber
220M and 600M sizes from one family, same interface
Strongest published open-source WER on Russian, Kazakh, Kyrgyz and Uzbek

Use Cases on Mixpeek

Transcribing Russian or Turkic-language audio before embedding the text

Multilingual call or meeting archives where Whisper underperforms on the language

Speech feature extraction from the ssl variants for audio similarity work

A second ASR pass to compare against an English-first model on mixed-language content

Benchmarks

Dataset	Metric	Score	Source
Common Voice, Russian	WER	7.1% (220M), 5.1% (600M), against 9.1% for Whisper large v3	GigaAM Multilingual model card, 2026
Internal in-the-wild, Kazakh	WER	18.8% (220M), 15.8% (600M), against 65.2% for Whisper large v3	GigaAM Multilingual model card, 2026
FLEURS, English	WER	12.2% (220M), 9.4% (600M), against 3.9% for Whisper large v3	GigaAM Multilingual model card, 2026 (the row where it loses)

Performance

Input SizeUtterances up to 30 seconds in the card's evaluation protocol

GPU LatencyInput dependent

GPU ThroughputBatch dependent

GPU MemoryModel dependent

Figures above are the model card's own, under greedy decoding with normalized references and digit-bearing references excluded. We have not reproduced them.

Common Pipeline Companions

Systran/faster-whisper-large-v3

The English side of a two-model ASR setup

intfloat/multilingual-e5-large-instruct

Embedding the transcript this model produces

Frequently Asked Questions

Is GigaAM Multilingual better than Whisper?

On Russian, Kazakh, Kyrgyz and Uzbek its card reports substantially lower WER, with Whisper large v3 above 60% on Kazakh and Kyrgyz in-the-wild sets where GigaAM is under 20%. On English, Whisper large v3 wins: 3.9% against 9.4% on FLEURS for the 600M variant. Route by language rather than picking one model for everything.

What is the difference between the ssl and ctc variants?

The ssl checkpoints are the self-supervised encoders, useful when you want speech representations rather than text. The ctc checkpoints add a character-wise CTC decoder and transcribe. Both come at 220M and 600M.

Which languages does it actually cover?

Pretraining spanned more than 70 languages, and the card names five in its metadata: Russian, English, Kazakh, Kyrgyz and Uzbek. Published WER exists for those five. Treat anything outside that list as untested.