IBM Releases Two Granite Speech 4.1 2B Models: High-Speed ASR and Translation

IBM Releases Two Granite Speech 4.1 2B Models: Autoregressive ASR with Translation and Non-Autoregressive Editing for Fast Inference

IBM has released the Granite Speech 4.1 2B and 2B-NAR models under the Apache 2.0 license to address the compute-accuracy trade-off in enterprise speech recognition. The standard model achieves a competitive mean Word Error Rate of 5.33 on the Open ASR Leaderboard as of April 2026.

Why This Matters

Enterprise AI teams frequently struggle with the technical reality that production-grade ASR systems typically demand massive compute resources or sacrifice transcription accuracy to maintain latency budgets. By optimizing a ~2B-parameter architecture, IBM demonstrates that careful modality adaptation and non-autoregressive editing can achieve high-fidelity results without the hardware overhead of larger models.

This release highlights the shift toward specialized, efficient models that can process audio at scale. The NAR variant’s ability to transcribe one hour of audio in under two seconds on a single H100 GPU provides a scalable path for real-time applications that previously required significant infrastructure investment.

Key Insights

• Granite Speech 4.1 2B scores a 1.33 WER on LibriSpeech clean and 5.33 mean WER on the Open ASR Leaderboard (2026).
• The 2B-NAR model achieves an RTFx of 1820 on a single H100 GPU using batched inference at batch size 128.
• The architecture features a 16-layer Conformer encoder trained with dual-head Connectionist Temporal Classification (CTC) for character and BPE units.
• A 2-layer window Q-Former downsamples acoustic embeddings by a factor of 10, resulting in a 10Hz embedding rate for the language model.
• The NAR variant utilizes a 1B-parameter bidirectional LLM editor based on Granite-4.0-1b-base with LoRA adaptation at rank 128.
• The standard autoregressive model supports six languages and bidirectional automatic speech translation (AST), whereas the NAR variant is limited to five languages for ASR only.

Practical Applications

• High-throughput transcription: Use Granite Speech 4.1 2B-NAR for large-scale archival processing where speed is critical. Pitfall: Attempting to use the NAR model for Japanese transcription or speech translation will result in failure as these features are exclusive to the autoregressive model.
• Meeting Intelligence: Deploy Granite Speech 4.1 2B-Plus for corporate environments requiring speaker-attributed ASR and word-level timestamps. Pitfall: Using the standard 2B model for multi-speaker logs will lack the necessary identity metadata and precise timing required for legal or compliance records.
• Multilingual Voice Assistants: Utilize the standard 2B model for bidirectional translation between English, French, German, Spanish, Portuguese, and Japanese. Pitfall: Neglecting to use flash_attention_2 for inference on the NAR model will prevent proper sequence packing and bidirectional context handling.

References:

• https://www.marktechpost.com/2026/04/30/ibm-releases-two-granite-speech-4-1-2b-models-autoregressive-asr-with-translation-and-non-autoregressive-editing-for-fast-inference/