Open-weight model · Speech recognition
nb-wav2vec2-1b-nynorsk
by Nasjonalbiblioteket AI Lab NbAiLab/nb-wav2vec2-1b-nynorsk
This model is finetuned on top of feature extractor XLS-R from Facebook/Meta. The finetuned model achieves the following results on the test set with a 5-gram KenLM.
Parameters963M
Context—
Weights11.7 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads1.2M
Runs On
What it takes to serve nb-wav2vec2-1b-nynorsk (963M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 1.9 GB | 2.3 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 1.0 GB | 1.2 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.5 GB | 0.6 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.
Model Card
By Nasjonalbiblioteket AI Lab, published under apache-2.0, revision 37107327f6af.
Norwegian Wav2Vec2 Model - 1B Nynorsk
This model is finetuned on top of feature extractor XLS-R from Facebook/Meta. The finetuned model achieves the following results on the test set with a 5-gram KenLM. The numbers in parentheses are the results without the language model: - WER: 0.1132 (0.1364) - CER: 0.0402 (---)
Model description
This is one of several Wav2Vec-models our team created during the hostedRobust Speech Event. This is the complete list of our models and their final scores:
| Model | Final WER | |
|---|---|---|
| NbAiLab/nb-wav2vec2-1b-bokmaal | 6.33 | |
| NbAiLab/nb-wav2vec2-300m-bokmaal | 7.03 | |
| NbAiLab/nb-wav2vec2-1b-nynorsk (this model) | 11.32 | |
| NbAiLab/nb-wav2vec2-300m-nynorsk | 12.22 |
Dataset
In parallel with the event, the team also converted the Norwegian Parliamentary Speech Corpus (NPSC) to the NbAiLab/NPSCin Dataset format and used that as the main source for training.
Code
Configuration
- Architecture
- Wav2Vec2ForCTC
- Layers
- 48
- Hidden size
- 1,280
- Feed-forward size
- 5,120
- Attention heads
- 16
- Vocabulary size
- 34
- Stored precision
- float32
- Model type
- wav2vec2
Identity and Version
- Repository
- NbAiLab/nb-wav2vec2-1b-nynorsk
- Publisher
- Nasjonalbiblioteket AI Lab
- Task
- Speech recognition
- Modality
- Audio
- Library
- transformers
- Parameters
- 963M parameters
- Languages
- nn, no
- Revision
- 37107327f6af742c008042d8cfdba4276de18f82
- First published
- 2022-06-09
- Last updated
- 2024-12-27
Files and Weights
36 files, 11.7 GB in total. The weights are 4 files totalling 11.7 GB in bin, safetensors.
Weights4 files · 11.7 GB
Configuration14 files · 128.9 KB
Tokenizer2 files · 571 B
Documentation2 files · 12.5 KB
Other12 files · 15.2 MB
Repository2 files · 1.3 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| language_model/5gram.bin | Weights | 4.0 GB | ab572020f487 |
| model.safetensors | Weights | 3.9 GB | 55b79d680168 |
| pytorch_model.bin | Weights | 3.9 GB | 5b5bc883abdd |
| training_args.bin | Weights | 3.1 KB | 1790380ec4ed |
| add_kenlm.py | Configuration | 1.3 KB | — |
| added_tokens.json | Configuration | 23 B | — |
| all_results.json | Configuration | 398 B | — |
| alphabet.json | Configuration | 223 B | — |
| cardinal_numbers.py | Configuration | 27.3 KB | — |
| config.json | Configuration | 2.0 KB | — |
| eval.py | Configuration | 10.8 KB | — |
| eval_results.json | Configuration | 223 B | — |
| language_model/attrs.json | Configuration | 78 B | — |
| preprocessor_config.json | Configuration | 260 B | — |
| run_speech_recognition_ctc.py | Configuration | 34.0 KB | — |
| special_tokens_map.json | Configuration | 695 B | — |
| train_results.json | Configuration | 196 B | — |
| trainer_state.json | Configuration | 51.4 KB | — |
| README.md | Documentation | 7.2 KB | — |
| README.old.md | Documentation | 5.4 KB | — |
| NbAiLab_NPSC_16K_mp3_nynorsk_test_eval_results.txt | Other | 50 B | — |
| NbAiLab_NPSC_16K_mp3_nynorsk_test_eval_results_no_lang_model.txt | Other | 47 B | — |
| language_model/unigrams.txt | Other | 14.9 MB | 48aeb02d80e4 |
| log_NbAiLab_NPSC_16K_mp3_nynorsk_test_predictions.txt | Other | 67.4 KB | — |
| log_NbAiLab_NPSC_16K_mp3_nynorsk_test_predictions_no_lang_model.txt | Other | 67.4 KB | — |
| log_NbAiLab_NPSC_16K_mp3_nynorsk_test_targets.txt | Other | 68.5 KB | — |
| log_NbAiLab_NPSC_16K_mp3_nynorsk_test_targets_no_lang_model.txt | Other | 68.5 KB | — |
| run.sh | Other | 1.5 KB | — |
| run_recover.sh | Other | 1.5 KB | — |
| runs/Mar24_16-58-45_dante/1648137570.9585772/events.out.tfevents.1648137570.dante.2925078.1 | Other | 4.8 KB | 17d568c75773 |
| runs/Mar24_16-58-45_dante/events.out.tfevents.1648137570.dante.2925078.0 | Other | 71.3 KB | 04228cc29243 |
| runs/Mar24_16-58-45_dante/events.out.tfevents.1648222995.dante.2925078.2 | Other | 364 B | 4b12aa934583 |
| .gitattributes | Repository | 1.3 KB | — |
| .gitignore | Repository | 21 B | — |
| tokenizer_config.json | Tokenizer | 282 B | — |
| vocab.json | Tokenizer | 289 B | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 11.7 GB
Download from Nasjonalbiblioteket AI Lab
Released by Nasjonalbiblioteket AI Lab through its official repository on Hugging Face. Read the license.
Built From
- Described by arXiv:2307.01672
- Trained on (disclosed) NbAiLab/NPSC
Evaluations
Each result is shown as reported, with the conditions its reporter stated. None is a SAVRN measurement. A comparison lines two results up only when their configuration, unit and setup are all stated and identical.
| Benchmark | Conditions | Result | Reported by | Revision | Date |
|---|---|---|---|---|---|
| NPSC | Task Automatic Speech RecognitionMetric Test (Nynorsk) CERComparison conditions not established | 0.0402637 | NbAiLab Publisher reported |
Evaluated revision not stated | — |
| NPSC | Task Automatic Speech RecognitionMetric Test (Nynorsk) WERComparison conditions not established | 0.113197 | NbAiLab Publisher reported |
Evaluated revision not stated | — |
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 11.7 GB |
| 16-bit | 1.9 GB |
| 8-bit | 1.0 GB |
| 4-bit | 0.5 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Compare nb-wav2vec2-1b-nynorsk
Questions About nb-wav2vec2-1b-nynorsk
How much GPU memory does nb-wav2vec2-1b-nynorsk need?
About 2.3 GB at 16-bit and 0.6 GB at 4-bit: the weights (963M parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run nb-wav2vec2-1b-nynorsk on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use nb-wav2vec2-1b-nynorsk commercially?
Yes. nb-wav2vec2-1b-nynorsk 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
parakeet-ctc-1.1b is an ASR model that transcribes speech in lower case English alphabet. This model is jointly developed by NVIDIA NeMo and Suno.ai teams. It is an XXL version of FastConformer CTC [1] (around 1.1B parameters) model. See the model architecture section and NeMo documentation for complete architecture details. To train, fine-tune or play with the model you will need to install NVIDIA NeMo. We recommend you install it after you've installed latest PyTorch version. There are several ways to use this model. Choose the one that fits your needs. NeMo-Speech.cpp provides a lightweight native C++ runtime for local inference with this model. After installing the runtime: See the…
Whisper is a state-of-the-art model for automatic speech recognition (ASR) and speech translation, proposed in the paper Robust Speech Recognition via Large-Scale Weak Supervision by Alec Radford et al. from OpenAI. Trained on >5M hours of labeled data, Whisper demonstrates a strong ability to generalise to many datasets and domains in a zero-shot setting. Whisper large-v3-turbo is a finetuned version of a pruned Whisper large-v3. In other words, it's the exact same model, except that the number of decoding layers have reduced from 32 to 4. As a result, the model is way faster, at the expense of a minor quality degradation. You can find more details about it in this GitHub discussion.…
Fine-tuned openai/whisper-large-v3-turbo on Monsoon Hindi ASR data using LoRA (r=32, alpha=64). English capabilities preserved with <10% relative degradation across all benchmarks. LoRA's structural separation (frozen base + low-rank adapters) provides a natural safeguard against catastrophic forgetting.
For usage instructions follow openai/whisper-large-v3-turbo. Note for faster-whisper vocab changes make model.ismultilingual and suppresstokens wrong. Please adjust the code as required if you want to use this with faster-whisper. Turbo finetune with japanese tokenizer. Full finetune trained 2^19 steps, batch size 64. Smaller vocab with ~1.6x bytes/token allows faster speed with 4 layers vs 2 layer distil (10% larger decoder). Benchmarks. Short form slightly behind v0.2 (trained less?) but long form much better. Also trained for lyrics but untested. Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC)
Whisper finetune for Japanese focused on general/anime domains. For usage instructions follow openai/whisper-large-v3-turbo. Due to vocab changes ctranslate2>=4.7.1 required for faster-whisper. For inference engines with hardcoded vocab, the token embedding can be padded. Finetuned from turbo with pruned vocab, indices can be found in mapping.txt. Trained decoder only for 2^20 steps, batch size 64. Using a 45000 hour corpus (largest source is 17000 of filtered reazonspeech-all) with custom mixing ratio and augmentation to maintain long form performance and timestamps. Benchmarks. Competitive/SOTA on test sets, slightly better than 1.5B on short form, worse on long form. Also trained for…