granite-4.0-1b-speech

by ibm-granite

#1 Open ASR Leaderboard at 1B — edge-deployable multilingual transcription

120Kdl/month

1Bparams

HuggingFace Use in Pipeline

Identifiers

Model ID

ibm-granite/granite-4.0-1b-speech

Feature URI

mixpeek://transcription@v1/ibm_granite_40_1b_speech_v1

Overview

Granite 4.0 1B Speech is the smallest model to reach #1 on the HuggingFace Open ASR Leaderboard. At just 1B parameters, it achieves 1.42% WER on LibriSpeech Clean and 5.52% average WER across benchmarks, while running at 280x realtime factor on GPU.

It supports English and Japanese with keyword list biasing for domain-specific vocabulary. The compact size makes it ideal for edge deployment, serverless functions, and cost-sensitive pipelines where Whisper Large v3 (1.5B) is too heavy. On Mixpeek, it serves as the default transcription model for latency-sensitive and high-volume audio processing.

Architecture

Compact encoder-decoder (1B parameters) optimized for throughput. Supports keyword biasing via attention-based shallow fusion. English + Japanese language support.

Mixpeek SDK Integration

import { Mixpeek } from "mixpeek";

const mx = new Mixpeek({ apiKey: "API_KEY" });

await mx.collections.ingest({
  collection_id: "my-collection",
  source: { url: "https://example.com/podcast.mp3" },
  feature_extractors: [{
    name: "transcription",
    version: "v1",
    params: {
      model_id: "ibm-granite/granite-4.0-1b-speech"
    }
  }]
});