Search public pages, research tools, and SAVRN solutions.
Open-weight model · Speech recognition
by Hosung Park kresnik/wav2vec2-large-xlsr-korean
What it takes to serve wav2vec2-large-xlsr-korean (317M 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 | 0.6 GB | 0.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 0.3 GB | 0.4 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.2 GB | 0.2 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.
Korean audio comes in, Korean text goes out, and the whole thing fits in 0.8 GB of memory at 16-bit precision, with the weights themselves at 0.6 GB. That is a fraction of one MI300X, which at 192 GB and $1.85 an hour on demand is the cheapest host in our table, so share the card with other workloads rather than dedicate it. The download is larger than the run: 15 files totaling 5.03 GB, because the stored precision is float32.
Under Apache 2.0 you may run it in production, modify it and redistribute it, keeping the license notices intact. The publisher, Hosung Park, trained it on kresnik/zeroth_korean and reports a test character error rate of 1.78 and a word error rate of 4.74 on that same set, publisher-reported figures. Test on your own recordings, and note the checkpoint dates to March 2022 with a last update in July 2023.
By Hosung Park, published under apache-2.0, revision 629c9a3501c1.
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from datasets import load_dataset
import soundfile as sf
import torch
from jiwer import wer
processor = Wav2Vec2Processor.from_pretrained("kresnik/wav2vec2-large-xlsr-korean")
model = Wav2Vec2ForCTC.from_pretrained("kresnik/wav2vec2-large-xlsr-korean").to('cuda')
ds = load_dataset("kresnik/zeroth_korean", "clean")
test_ds = ds['test']
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
test_ds = test_ds.map(map_to_array)
def map_to_pred(batch):
inputs = processor(batch["speech"], sampling_rate=16000, return_tensors="pt", padding="longest")
input_values = inputs.input_values.to("cuda")
with torch.no_grad():
logits = model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)
batch["transcription"] = transcription
return batch
result = test_ds.map(map_to_pred, batched=True, batch_size=16, remove_columns=["speech"])
print("WER:", wer(result["text"], result["transcription"]))
15 files, 5.0 GB in total. The weights are 7 files totalling 5.0 GB in bin, pt, pth, safetensors.
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 1.3 GB | 29be111b9247 |
| optimizer.pt | Weights | 2.5 GB | df158ff3490a |
| pytorch_model.bin | Weights | 1.3 GB | 762e434b6f07 |
| rng_state.pth | Weights | 14.5 KB | edea941b5d49 |
| scaler.pt | Weights | 559 B | 58caf7038dad |
| scheduler.pt | Weights | 623 B | 18e7c2b053fc |
| training_args.bin | Weights | 2.9 KB | beb15ba46f1b |
| config.json | Configuration | 2.3 KB | — |
| preprocessor_config.json | Configuration | 214 B | — |
| trainer_state.json | Configuration | 511.6 KB | — |
| README.md | Documentation | 1.9 KB | — |
| special_tokens_map.config | Other | 85 B | — |
| .gitattributes | Repository | 1.2 KB | — |
| tokenizer_config.json | Tokenizer | 161 B | — |
| vocab.json | Tokenizer | 18.2 KB | — |
Released by Hosung Park through its official repository on Hugging Face. Read the license.
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 |
|---|---|---|---|---|---|
| Zeroth Korean | Task Automatic Speech RecognitionMetric Test CERComparison conditions not established | 1.78 | kresnik Publisher reported |
Evaluated revision not stated | — |
| Zeroth Korean | Task Automatic Speech RecognitionMetric Test WERComparison conditions not established | 4.74 | kresnik Publisher reported |
Evaluated revision not stated | — |
| Precision | Weights in memory |
|---|---|
| As published | 5.0 GB |
| 16-bit | 0.6 GB |
| 8-bit | 0.3 GB |
| 4-bit | 0.2 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
About 0.8 GB at 16-bit and 0.2 GB at 4-bit: the weights (317M parameters) plus a working margin. A long context needs more.
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.
Yes. wav2vec2-large-xlsr-korean 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.
Fine-tuned facebook/wav2vec2-large-xlsr-53 hindi using the Multilingual and code-switching ASR challenges for low resource Indian languages. When using this model, make sure that your speech input is sampled at 16kHz. The model can be used directly (without a language model) as follows: The model can be evaluated as follows on the hindi test data of Common Voice. The script used for training can be found Hindi ASR Fine Tuning Wav2Vec2
This model is a finetuned version of facebook/wav2vec2-large-xlsr-53. There was an issue with vocab, seems like there are special characters included and they were not considered during training You could try
A fine-tuned XLS-R 300M CTC model for Urdu automatic speech recognition. It transcribes 16 kHz mono audio and includes an optional 5-gram KenLM decoder. Best reported result: 39.89% WER / 16.70% CER with KenLM decoding on the Urdu Common Voice 8.0 test set. See the Kaggle evaluation notebook for a reproducible example. The repository contains a 5-gram KenLM language model. The Kaggle notebook evaluates a five-sample streaming smoke test from fixie-ai/commonvoice170 (ur, test). Results are reported on the Urdu test split of Mozilla Common Voice 8.0. The language-model row is the model-card score; compare each result only with the same decoding strategy. To reproduce language-model evaluation…
IPA phoneme recognition for Kölsch (Ripuarian German, Cologne). Fine-tuned from facebook/wav2vec2-xls-r-300m with a CTC head over a 48-symbol IPA inventory. This is also the forced-alignment model. Its frame-level CTC posteriors drive torchaudio.functional.forcedalign to produce Praat TextGrids with word and phone tiers — see notebook 9 in the repository. Trained during the CIF Tandem Fellowship at IfL-Phonetik, University of Cologne, on Alles Kölsch (Bhatt & Lindlar 1998) — 4,670 utterances, 4.5 hours of spontaneous Cologne dialect speech from 105 speakers aged 10–88 across 49 neighbourhoods. Kölsch is Ripuarian German. It has no public speech dataset and no standardised spelling: 94.1 %…
Model · Speech recognition
CLASSLA - CLARIN Knowledge Centre for South Slavic Languages
This model for Croatian ASR is based on the facebook/wav2vec2-xls-r-300m model and was fine-tuned with 300 hours of recordings and transcripts from the ASR Croatian parliament dataset ParlaSpeech-HR v1.0. If you use this model, please cite the following paper: Nikola Ljubešić, Danijel Koržinek, Peter Rupnik, Ivo-Pavao Jazbec. ParlaSpeech-HR -- a freely available ASR dataset for Croatian bootstrapped from the ParlaMint corpus. http://www.lrec-conf.org/proceedings/lrec2022/workshops/ParlaCLARINIII/pdf/2022.parlaclariniii-1.16.pdf Evaluation is performed on the dev and test portions of the ParlaSpeech-HR v1.0 dataset. There are multiple models available, and in terms of CER and WER, the…
Finetuned version of KBs VoxRex large model using Swedish radio broadcasts, NST and Common Voice data. Evalutation without a language model gives the following: WER for NST + Common Voice test set (2% of total sentences) is 2.5%. WER for Common Voice test set is 8.49% directly and 7.37% with a 4-gram language model. When using this model, make sure that your speech input is sampled at 16kHz. This model has been fine-tuned for 120000 updates on NST + CommonVoice and then for an additional 20000 updates on CommonVoice only. The additional fine-tuning on CommonVoice hurts performance on the NST+CommonVoice test set somewhat and, unsurprisingly, improves it on the CommonVoice test set. It seems…