Ported from upstream commit
99a4df9,
pinned 2026-05-24.
Validated against the Transformers reference at transcribe.cpp commit
c53af2c
on 2026-05-24.
Offline multilingual speech-to-text in a single non-autoregressive editor
pass. IBM Granite Speech 4.1-2b NAR shares the Conformer audio encoder
with the AR Granite-Speech family but pairs it with a custom MLP-with-
attention projector and the Granite-4.0-1b LLM used as a bidirectional
editor (causal mask disabled). One forward pass produces logits over the
full transcript; CTC decode yields the final text. No token-by-token loop.
Takes a 16 kHz mono WAV and produces a transcript. English plus French,
German, Spanish, and Portuguese; ASR only (no translation, no
timestamps).
WER measured on the full LibriSpeech test-clean split (2620 utterances).
BF16 reference baseline (transformers model.transcribe, MPS, re-run
locally): 1.28% — matches the upstream model card's 1.29% to within
sampling noise. Text normalizer: Whisper EnglishTextNormalizer, the
same normalizer Open ASR Leaderboard uses. Reference reproduction
follows the model card path verbatim (AutoProcessor +
AutoModel.transcribe + processor.batch_decode) at HF revision
99a4df9 (single-file modeling_granite_speech_nar.py snapshot, the
README's canonical target); no mask patching is required because the
NAR LM uses create_bidirectional_mask() natively. F16, Q8_0, and
Q6_K all match BF16's 1.29%; Q5_K_M dips slightly to 1.25% (within
overlapping CIs).
If your audio isn't already 16 kHz mono WAV, convert it first:
ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav
See the transcribe.cpp model page for performance
numbers, numerical validation, and reproduction steps.
License
Inherited from the base model: Apache-2.0. See the
upstream model card for full terms.
Original Model Card
The section below is reproduced from
ibm-granite/granite-speech-4.1-2b-nar at commit
99a4df9 for offline reference. The upstream card is the
authoritative source.
Granite-Speech-4.1-2B-NAR
Model Summary:
Granite-Speech-4.1-2B-NAR is a non-autoregressive (NAR) speech recognition model that formulates ASR as conditional transcript editing.
Instead of decoding tokens one at a time, it edits a CTC hypothesis in a single forward pass using a bidirectional LLM, achieving competitive accuracy with faster inference than autoregressive alternatives.
The model is based on the NLE (Non-autoregressive LLM-based Editing) architecture described in this paper.
For applications where accuracy is the primary concern, consider granite-speech-4.1-2b, an autoregressive model from the Granite Speech 4.1 family which achieves higher transcription accuracy at the cost of increased inference latency. Granite-speech-4.1-2b produces punctuated and capitalized transcripts, supports AST and keyword-biased recognition, and includes Japanese.
When speaker or word-timing information is needed, consider using granite-speech-4.1-2b-plus, which extends the above model with speaker-attributed ASR (speaker labels + word transcripts) and word-level timing information.
Intended Use:
The model is intended for automatic speech recognition tasks, particularly in latency-sensitive applications where fast inference is critical.
Greedy decoding with bfloat16 inference. WER computed with jiwer after whisper_normalizerEnglishTextNormalizer normalization.
Open ASR Leaderboard results may differ slightly due to normalization and scoring pipeline differences.
Measured RTFx of ~1820 on a single H100 GPU (batched inference, batch size 128).
Dataset
WER
Dataset
WER
LibriSpeech clean
1.29
MLS EN
4.77
LibriSpeech other
2.75
MLS DE
4.75
CommonVoice 15 EN
6.50
MLS ES
3.31
CommonVoice 15 DE
4.73
MLS FR
4.52
CommonVoice 15 ES
4.02
MLS PT
11.86
CommonVoice 15 FR
7.17
AMI IHM
7.91
CommonVoice 15 PT
2.57
AMI SDM
19.59
Earnings-22
8.48
GigaSpeech
10.12
SPGISpeech
3.04
TED-LIUM
3.67
VoxPopuli
5.83
Usage
Installation
We require flash_attention_2 for inference, since this backend supports sequence packing and respects the is_causal=False flag.
Requires transformers>=5.5.3 and torch>=2.9.1.
A 16-layer Conformer encoder trained with CTC on character-level targets. It processes 16kHz audio with stacked log-mel features (80 mel bins, 2-frame stacking) and uses block attention with 4-second audio blocks and self-conditioning at layer 8.
The encoder has a dual CTC head: alongside the character-level output, a secondary BPE head produces CTC logits over the LLM's 100K token vocabulary. The BPE head uses posterior-weighted pooling (window size 4) with importance weights derived from mid-layer blank probabilities (1 - blank_prob).
(2) Q-Former Projector (160M params)
A 2-layer window Q-Former that downsamples the concatenated hidden representations from 4 encoder layers (layers 4, 8, 12, 16) by 5x.
Each 15-frame window is reduced to 3 queries via cross-attention, resulting in a 10Hz acoustic embedding rate for the LLM (2x from encoder + 5x from projector).
granite-4.0-1b-base with its causal attention mask removed, enabling bidirectional context.
Adapted with LoRA (rank 128) applied to both attention and MLP layers. The LLM receives concatenated audio embeddings and an interleaved CTC hypothesis with insertion slots, then predicts the edited transcript in a single parallel forward pass using a CTC objective.
How Granite-speech NAR Works
Architecture
The frozen CTC encoder produces acoustic embeddings and an initial hypothesis.
The hypothesis is interleaved with insertion slots (blank tokens between each token)
The projected audio embeddings are concatenated with the interleaved hypothesis embeddings
The bidirectional LLM predicts edits (copy, insert, delete, replace) at all positions simultaneously
CTC greedy decoding (argmax + collapse) produces the final transcript
This design exploits the identity mapping bias of Transformers: residual connections and tied embeddings make the model naturally inclined to copy input tokens, so it focuses learning capacity on corrections rather than full reconstruction.
Training Data
The model was trained on approximately 130K hours of speech across five languages (English, Spanish, French, German, Portuguese), using publicly available datasets including CommonVoice 15, MLS, LibriSpeech, Libriheavy long, AMI, Granary VoxPopuli, Granary YODAS, Earnings-22, Fisher, CallHome, and SwitchBoard.
For additional training details, see the paper.
Infrastructure
Training was completed on IBM's Blue Vela cluster using 16 H100 GPUs (2 nodes) for 5 epochs (3 days).
Ethical Considerations and Limitations
The model is designed specifically for automatic speech recognition and does not generate free-form text, which limits the risk of hallucination compared to general-purpose speech-language models.
However, transcription accuracy varies across languages and acoustic conditions. Performance may be weaker on languages with less training data (e.g., Portuguese) or in challenging acoustic environments (e.g., far-field, overlapping speech).
The model's editing approach is conservative by design — it prefers deletions over insertions, which reduces hallucination risk but may occasionally drop words in noisy conditions.
To enhance safety, we recommend using granite-speech-4.1-2b-nar alongside Granite Guardian. Granite Guardian is a fine-tuned instruct model designed to detect and flag risks in prompts and responses across key dimensions outlined in the IBM AI Risk Atlas.
