granite-speech-4.1-2b-nar-glade by Josh Leverette

Model Card

By Josh Leverette, published under apache-2.0, revision 350774a2193b.

Glade is a native Swift library for on-device speech inference on Apple silicon. This repository supplies converted model assets and configuration for Glade's model-specific host runtime. Glade's runtime source is currently private. A community conversion of IBM Granite Speech 4.1 2B NAR for Glade, using Apple's Core AI APIs and optimized for Neural Engine execution on Apple silicon. The model combines a CTC acoustic encoder with one bidirectional, audio-conditioned transcript-editing pass. This is a W8A16 export: most convolution/projection weights are INT8 and activations are FP16. Conformer convolution output projections, normalization parameters, and the embedding table remain FP16.…

Read Josh Leverette's full model card

Granite Speech 4.1 2B NAR for Glade

Glade is a native Swift library for on-device speech inference on Apple silicon. This repository supplies converted model assets and configuration for Glade's model-specific host runtime. Glade's runtime source is currently private.

A community conversion of IBM Granite Speech 4.1 2B NAR for Glade, using Apple's Core AI APIs and optimized for Neural Engine execution on Apple silicon. The model combines a CTC acoustic encoder with one bidirectional, audio-conditioned transcript-editing pass.

This is a W8A16 export: most convolution/projection weights are INT8 and activations are FP16. Conformer convolution output projections, normalization parameters, and the embedding table remain FP16. Graphs use ANE-oriented layouts, fused layers, and shared weights across a bounded set of static input shapes. Debug source metadata has been removed.

The repository contains 59 .aimodel components, embeddings.f16, tokenizer.json, model configuration in metadata.json. The components cover the encoder, acoustic projector, CTC head, editor, and final vocabulary head. These are portable source assets; Core AI specializes them for the target device on first use, which can take several minutes.

Requirements: Apple silicon with macOS 27 or a physical iOS 27 device supporting Core AI. These assets require application-side orchestration, audio preprocessing, CTC collapse, and tokenization; they are not a drop-in Transformers checkpoint. No inference runtime is included here. Use 16 kHz mono audio and bounded chunks (60 seconds or shorter) for longer recordings. The model is intended for final transcription, rather than native streaming.

Converted from upstream revision a1e3416e25ce29ab3852778e54fa8b3bd59c4bf2. Original model by IBM; this community conversion is not an official IBM or Apple release. Distributed under Apache-2.0, matching the upstream model. See the upstream model card for training details and limitations.

Measured performance

M3 MacBook Air (16 GB), macOS 27. The audio is JFK's “We choose to go to the Moon” speech: a 20-second excerpt and the full 18-minute 15-second recording. The qualified host uses shared encoder/editor shapes and bounded chunks of at most 60 seconds, with CPU Silero choosing boundaries. Silero is not included in this repository.

Speech excerpt	Warm transcription	Audio / elapsed
20 seconds	0.408 s	49.0×
18 min 15 s	21.90 s	50.0×

Medians of three runs after preparation and warmup, including audio reading, chunk planning, the frontend, encoder, projector, editor and decoding. Preparation is excluded. Both measurements used macOS thermal state “fair” after compilation. The full recording retains all samples across 21 chunks. Benchmark-normalized WER is 52/2,219 (2.34%) against the supplied reference; this is one recording, not a general or multilingual accuracy evaluation. A separate full-recording trace recorded 1,844 ANE predictions and zero GPU intervals. Placement is not a measure of arithmetic utilization.

Device specialization and loading

For the same 59-component shared-shape source configuration on this Mac:

Observed preparation	Time
First full-model preparation, including recovery of missing editor caches	24 min 2 s
Immediate cached relaunch	0.775 s
Later preparation after further cache-load failures	19 min 6 s
Following cached short-clip launch	0.762 s

Preparation includes model loading and compilation requested by Core AI, before warmup or transcription. These observations do not isolate a clean first-install compile or the incremental cost of adding shapes. They preceded removal of authoring debug metadata, which preserves computation but changes artifact identity. Underlying ANE cache state is not fully observable, and cache loss or OS/device/application changes can require specialization again. These Mac measurements are not iPhone startup estimates.

Identity and Version

Repository: coder543/granite-speech-4.1-2b-nar-glade
Publisher: Josh Leverette
Task: Speech recognition
Modality: Audio
Library: Not stated by the source
Parameters: Not stated by the source
Languages: en, fr, de, es, pt
Revision: 350774a2193b17a079c17557aa2b259d746fa621
First published: 2026-09-26
Last updated: 2026-10-09

Files and Weights

183 files, 2.7 GB in total.

Configuration60 files · 10.0 KB

Tokenizer1 file · 7.2 MB

Documentation2 files · 15.9 KB

Other119 files · 2.7 GB

Repository1 file · 5.4 KB

Every file

File	Type	Size	SHA-256
ctc_head.aimodel/metadata.json	Configuration	105 B	—
editor_0.aimodel/metadata.json	Configuration	105 B	—
editor_1.aimodel/metadata.json	Configuration	105 B	—
editor_10.aimodel/metadata.json	Configuration	105 B	—
editor_11.aimodel/metadata.json	Configuration	105 B	—
editor_12.aimodel/metadata.json	Configuration	105 B	—
editor_13.aimodel/metadata.json	Configuration	105 B	—
editor_14.aimodel/metadata.json	Configuration	105 B	—
editor_15.aimodel/metadata.json	Configuration	105 B	—
editor_16.aimodel/metadata.json	Configuration	105 B	—
editor_17.aimodel/metadata.json	Configuration	105 B	—
editor_18.aimodel/metadata.json	Configuration	105 B	—
editor_19.aimodel/metadata.json	Configuration	105 B	—
editor_2.aimodel/metadata.json	Configuration	105 B	—
editor_20.aimodel/metadata.json	Configuration	105 B	—
editor_21.aimodel/metadata.json	Configuration	105 B	—
editor_22.aimodel/metadata.json	Configuration	105 B	—
editor_23.aimodel/metadata.json	Configuration	105 B	—
editor_24.aimodel/metadata.json	Configuration	105 B	—
editor_25.aimodel/metadata.json	Configuration	105 B	—
editor_26.aimodel/metadata.json	Configuration	105 B	—
editor_27.aimodel/metadata.json	Configuration	105 B	—
editor_28.aimodel/metadata.json	Configuration	105 B	—
editor_29.aimodel/metadata.json	Configuration	105 B	—
editor_3.aimodel/metadata.json	Configuration	105 B	—
editor_30.aimodel/metadata.json	Configuration	105 B	—
editor_31.aimodel/metadata.json	Configuration	105 B	—
editor_32.aimodel/metadata.json	Configuration	105 B	—
editor_33.aimodel/metadata.json	Configuration	105 B	—
editor_34.aimodel/metadata.json	Configuration	105 B	—
editor_35.aimodel/metadata.json	Configuration	105 B	—
editor_36.aimodel/metadata.json	Configuration	105 B	—
editor_37.aimodel/metadata.json	Configuration	105 B	—
editor_38.aimodel/metadata.json	Configuration	105 B	—
editor_39.aimodel/metadata.json	Configuration	105 B	—
editor_4.aimodel/metadata.json	Configuration	105 B	—
editor_5.aimodel/metadata.json	Configuration	105 B	—
editor_6.aimodel/metadata.json	Configuration	105 B	—
editor_7.aimodel/metadata.json	Configuration	105 B	—
editor_8.aimodel/metadata.json	Configuration	105 B	—
editor_9.aimodel/metadata.json	Configuration	105 B	—
editor_head.aimodel/metadata.json	Configuration	105 B	—
encoder_0.aimodel/metadata.json	Configuration	105 B	—
encoder_1.aimodel/metadata.json	Configuration	105 B	—
encoder_10.aimodel/metadata.json	Configuration	105 B	—
encoder_11.aimodel/metadata.json	Configuration	105 B	—
encoder_12.aimodel/metadata.json	Configuration	105 B	—
encoder_13.aimodel/metadata.json	Configuration	105 B	—
encoder_14.aimodel/metadata.json	Configuration	105 B	—
encoder_15.aimodel/metadata.json	Configuration	105 B	—
encoder_2.aimodel/metadata.json	Configuration	105 B	—
encoder_3.aimodel/metadata.json	Configuration	105 B	—
encoder_4.aimodel/metadata.json	Configuration	105 B	—
encoder_5.aimodel/metadata.json	Configuration	105 B	—
encoder_6.aimodel/metadata.json	Configuration	105 B	—
encoder_7.aimodel/metadata.json	Configuration	105 B	—
encoder_8.aimodel/metadata.json	Configuration	105 B	—
encoder_9.aimodel/metadata.json	Configuration	105 B	—
metadata.json	Configuration	3.8 KB	—
projector.aimodel/metadata.json	Configuration	105 B	—
LICENSE	Documentation	11.4 KB	—
README.md	Documentation	4.5 KB	—
ctc_head.aimodel/main.hash	Other	32 B	—
ctc_head.aimodel/main.mlirb	Other	103.2 MB	077a65f76b54
editor_0.aimodel/main.hash	Other	32 B	—
editor_0.aimodel/main.mlirb	Other	35.8 MB	e0ee6d6fe226
editor_1.aimodel/main.hash	Other	32 B	—
editor_1.aimodel/main.mlirb	Other	35.8 MB	6cf96f4df2d9
editor_10.aimodel/main.hash	Other	32 B	—
editor_10.aimodel/main.mlirb	Other	35.8 MB	90b936f6b1e7
editor_11.aimodel/main.hash	Other	32 B	—
editor_11.aimodel/main.mlirb	Other	35.8 MB	77567f49d0ff
editor_12.aimodel/main.hash	Other	32 B	—
editor_12.aimodel/main.mlirb	Other	35.8 MB	9402c1ff6f49
editor_13.aimodel/main.hash	Other	32 B	—
editor_13.aimodel/main.mlirb	Other	35.8 MB	bc46e6be16e3
editor_14.aimodel/main.hash	Other	32 B	—
editor_14.aimodel/main.mlirb	Other	35.8 MB	78d90311f227
editor_15.aimodel/main.hash	Other	32 B	—
editor_15.aimodel/main.mlirb	Other	35.8 MB	1e3728cf518d
editor_16.aimodel/main.hash	Other	32 B	—
editor_16.aimodel/main.mlirb	Other	35.8 MB	8f444af96bd1
editor_17.aimodel/main.hash	Other	32 B	—
editor_17.aimodel/main.mlirb	Other	35.8 MB	61b01fa127c5
editor_18.aimodel/main.hash	Other	32 B	—
editor_18.aimodel/main.mlirb	Other	35.8 MB	1d00b4439d03
editor_19.aimodel/main.hash	Other	32 B	—
editor_19.aimodel/main.mlirb	Other	35.8 MB	6eeb5d14adf3
editor_2.aimodel/main.hash	Other	32 B	—
editor_2.aimodel/main.mlirb	Other	35.8 MB	01ed44afa3f0
editor_20.aimodel/main.hash	Other	32 B	—
editor_20.aimodel/main.mlirb	Other	35.8 MB	15854cbdaecf
editor_21.aimodel/main.hash	Other	32 B	—
editor_21.aimodel/main.mlirb	Other	35.8 MB	a3c353c4be45
editor_22.aimodel/main.hash	Other	32 B	—
editor_22.aimodel/main.mlirb	Other	35.8 MB	54c1446f9c59
editor_23.aimodel/main.hash	Other	32 B	—
editor_23.aimodel/main.mlirb	Other	35.8 MB	7dcbe692324e
editor_24.aimodel/main.hash	Other	32 B	—
editor_24.aimodel/main.mlirb	Other	35.8 MB	c041bfa7ea85
editor_25.aimodel/main.hash	Other	32 B	—
editor_25.aimodel/main.mlirb	Other	35.8 MB	5ecbc49fb602
editor_26.aimodel/main.hash	Other	32 B	—
editor_26.aimodel/main.mlirb	Other	35.8 MB	34944dca810b
editor_27.aimodel/main.hash	Other	32 B	—
editor_27.aimodel/main.mlirb	Other	35.8 MB	2a43cd95a335
editor_28.aimodel/main.hash	Other	32 B	—
editor_28.aimodel/main.mlirb	Other	35.8 MB	33e51422c929
editor_29.aimodel/main.hash	Other	32 B	—
editor_29.aimodel/main.mlirb	Other	35.8 MB	4eca1c85bdf2
editor_3.aimodel/main.hash	Other	32 B	—
editor_3.aimodel/main.mlirb	Other	35.8 MB	92eef69cd196
editor_30.aimodel/main.hash	Other	32 B	—
editor_30.aimodel/main.mlirb	Other	35.8 MB	40749199ae1d
editor_31.aimodel/main.hash	Other	32 B	—
editor_31.aimodel/main.mlirb	Other	35.8 MB	6f8b5235e783
editor_32.aimodel/main.hash	Other	32 B	—
editor_32.aimodel/main.mlirb	Other	35.8 MB	712a135913ea
editor_33.aimodel/main.hash	Other	32 B	—
editor_33.aimodel/main.mlirb	Other	35.8 MB	87edab34fb43
editor_34.aimodel/main.hash	Other	32 B	—
editor_34.aimodel/main.mlirb	Other	35.8 MB	e0ff5e8ed441
editor_35.aimodel/main.hash	Other	32 B	—
editor_35.aimodel/main.mlirb	Other	35.8 MB	5ec7f808af85
editor_36.aimodel/main.hash	Other	32 B	—
editor_36.aimodel/main.mlirb	Other	35.8 MB	58e4ffb83096
editor_37.aimodel/main.hash	Other	32 B	—
editor_37.aimodel/main.mlirb	Other	35.8 MB	0aa323aa8c4c
editor_38.aimodel/main.hash	Other	32 B	—
editor_38.aimodel/main.mlirb	Other	35.8 MB	1b19bdcb85e8
editor_39.aimodel/main.hash	Other	32 B	—
editor_39.aimodel/main.mlirb	Other	35.8 MB	4f676b395c08
editor_4.aimodel/main.hash	Other	32 B	—
editor_4.aimodel/main.mlirb	Other	35.8 MB	33dc4e326cd1
editor_5.aimodel/main.hash	Other	32 B	—
editor_5.aimodel/main.mlirb	Other	35.8 MB	fa668ac8ad5b
editor_6.aimodel/main.hash	Other	32 B	—
editor_6.aimodel/main.mlirb	Other	35.8 MB	c322afe094c3
editor_7.aimodel/main.hash	Other	32 B	—
editor_7.aimodel/main.mlirb	Other	35.8 MB	f489dc4b73cf
editor_8.aimodel/main.hash	Other	32 B	—
editor_8.aimodel/main.mlirb	Other	35.8 MB	fdfb9eba312d
editor_9.aimodel/main.hash	Other	32 B	—
editor_9.aimodel/main.mlirb	Other	35.8 MB	a2d71a320e3a
editor_head.aimodel/main.hash	Other	32 B	—
editor_head.aimodel/main.mlirb	Other	205.8 MB	00c29cd8ecf5
embeddings.f16	Other	411.0 MB	1fa3d6f7fea9
encoder_0.aimodel/main.hash	Other	32 B	—
encoder_0.aimodel/main.mlirb	Other	30.9 MB	331382243b6a
encoder_1.aimodel/main.hash	Other	32 B	—
encoder_1.aimodel/main.mlirb	Other	30.7 MB	7aeacbe4f89f
encoder_10.aimodel/main.hash	Other	32 B	—
encoder_10.aimodel/main.mlirb	Other	30.7 MB	a2770784610f
encoder_11.aimodel/main.hash	Other	32 B	—
encoder_11.aimodel/main.mlirb	Other	30.7 MB	dcfcc0743237
encoder_12.aimodel/main.hash	Other	32 B	—
encoder_12.aimodel/main.mlirb	Other	30.7 MB	b81e47c9a2ef
encoder_13.aimodel/main.hash	Other	32 B	—
encoder_13.aimodel/main.mlirb	Other	30.7 MB	1dd7bb35542c
encoder_14.aimodel/main.hash	Other	32 B	—
encoder_14.aimodel/main.mlirb	Other	30.7 MB	f394acf42d39
encoder_15.aimodel/main.hash	Other	32 B	—
encoder_15.aimodel/main.mlirb	Other	30.7 MB	e3e96e140e41
encoder_2.aimodel/main.hash	Other	32 B	—
encoder_2.aimodel/main.mlirb	Other	30.7 MB	b0d528033a96
encoder_3.aimodel/main.hash	Other	32 B	—
encoder_3.aimodel/main.mlirb	Other	30.7 MB	6ace1e70898e
encoder_4.aimodel/main.hash	Other	32 B	—
encoder_4.aimodel/main.mlirb	Other	30.7 MB	c6e75b8b7f1a
encoder_5.aimodel/main.hash	Other	32 B	—
encoder_5.aimodel/main.mlirb	Other	30.7 MB	288cb2a99998
encoder_6.aimodel/main.hash	Other	32 B	—
encoder_6.aimodel/main.mlirb	Other	30.7 MB	c27971760a84
encoder_7.aimodel/main.hash	Other	32 B	—
encoder_7.aimodel/main.mlirb	Other	31.4 MB	710d1836d5a3
encoder_8.aimodel/main.hash	Other	32 B	—
encoder_8.aimodel/main.mlirb	Other	30.7 MB	4906a0111d81
encoder_9.aimodel/main.hash	Other	32 B	—
encoder_9.aimodel/main.mlirb	Other	30.7 MB	bff943939e2b
projector.aimodel/main.hash	Other	32 B	—
projector.aimodel/main.mlirb	Other	80.0 MB	44b3484d4a41
.gitattributes	Repository	5.4 KB	—
tokenizer.json	Tokenizer	7.2 MB	—

License and Download

License: apache-2.0
Access: Open weights, no gate

Download from Josh Leverette

Released by Josh Leverette through its official repository on Hugging Face. Read the license.

Built From

Derived from ibm-granite/granite-speech-4.1-2b-nar
Quantized from ibm-granite/granite-speech-4.1-2b-nar

Questions About granite-speech-4.1-2b-nar-glade

Can I use granite-speech-4.1-2b-nar-glade commercially?

Yes. granite-speech-4.1-2b-nar-glade is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Model · Speech recognition

wav2vec2-large-xlsr-53-japanese

Jonatas Grosman

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Japanese using the train and validation splits of Common Voice 6.1, CSS10 and JSUT. When using this model, make sure that your speech input is sampled at 16kHz. This model has been fine-tuned thanks to the GPU credits generously given by the OVHcloud:) The script used for training can be found here: https://github.com/jonatasgrosman/wav2vec2-sprint The model can be used directly (without a language model) as follows... Using the HuggingSound library: The model can be evaluated as follows on the Japanese test data of Common Voice. In the table below I report the Word Error Rate (WER) and the Character Error Rate (CER) of the model. I ran the…

Open weights apache-2.0 transformers

View model

Model · Speech recognition

whisperkit-coreml

Argmax

WhisperKit is part of Argmax OSS, an On-device Speech AI SDK for Apple Silicon: https://github.com/argmaxinc/argmax-oss-swift Check out the WhisperKit paper and presentation from ICML 2025: https://icml.cc/virtual/2025/47854 For real-time transcription with speakers and custom vocabulary, check out Argmax Pro SDK: https://www.argmaxinc.com/blog/argmax-sdk-2

Open weights mit whisperkit

View model

Model · Speech recognition

speaker-diarization-3.1

Pyannote

Using this open-source model in production? Consider switching to pyannoteAI for better and faster options. This pipeline is the same as pyannote/speaker-diarization-3.0 except it removes the problematic use of onnxruntime. Both speaker segmentation and embedding now run in pure PyTorch. This should ease deployment and possibly speed up inference. It requires pyannote.audio version 3.1 or higher. It ingests mono audio sampled at 16kHz and outputs speaker diarization as an Annotation instance: - stereo or multi-channel audio files are automatically downmixed to mono by averaging the channels. - audio files sampled at a different rate are resampled to 16kHz automatically upon loading. 1.…

Access requested at publisher mit pyannote-audio

View model

Model · Speech recognition

speaker-diarization-community-1

Pyannote

This pipeline ingests mono audio sampled at 16kHz and outputs speaker diarization. - stereo or multi-channel audio files are automatically downmixed to mono by averaging the channels. - audio files sampled at a different rate are resampled to 16kHz automatically upon loading. The main improvements brought by Community-1 are: - improved speaker assignment and counting - simpler reconciliation with transcription timestamps with exclusive speaker diarization - easy offline use (i.e. without internet connection) - (optionally) hosted on pyannoteAI cloud 1. pip install pyannote.audio 3. Create access token at hf.co/settings/tokens. Out of the box, Community-1 is much better than…

Access requested at publisher cc-by-4.0 pyannote-audio

View model

Model · Speech recognition

faster-whisper-small

Systran

This repository contains the conversion of openai/whisper-small to the CTranslate2 model format. This model can be used in CTranslate2 or projects based on CTranslate2 such as faster-whisper. The original model was converted with the following command: Note that the model weights are saved in FP16. This type can be changed when the model is loaded using the computetype option in CTranslate2. For more information about the original model, see its model card.

Open weights mit ctranslate2

View model

Model · Speech recognition

wav2vec2-large-xlsr-53-portuguese

Jonatas Grosman

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Portuguese using the train and validation splits of Common Voice 6.1. When using this model, make sure that your speech input is sampled at 16kHz. This model has been fine-tuned thanks to the GPU credits generously given by the OVHcloud:) The script used for training can be found here: https://github.com/jonatasgrosman/wav2vec2-sprint The model can be used directly (without a language model) as follows... Using the HuggingSound library: 1. To evaluate on mozilla-foundation/commonvoice60 with split test 2. To evaluate on speech-recognition-community-v2/devdata If you want to cite this model you can use this

Open weights apache-2.0 transformers

View model