Search public pages, research tools, and SAVRN solutions.
Open-weight model · Speech recognition
by Vladimir Gurevich imvladikon/wav2vec2-xls-r-300m-hebrew
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the private datasets in 2 stages - firstly was fine-tuned on a small dataset with good samples Then the obtained model was fine-tuned on a large dataset with the small good dataset, with…
What it takes to serve wav2vec2-xls-r-300m-hebrew (315M 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.
A vocabulary of 32 tokens tells you this is not a language model. It is a CTC speech recognizer for Hebrew that Vladimir Gurevich fine-tuned from facebook/wav2vec2-xls-r-300m in two stages, the second on a large mixed set that included weakly labeled audio. At 315M parameters it runs in 0.8 GB at 16-bit, 0.4 GB at 8-bit and 0.2 GB at 4-bit; the cheapest slot is one MI300X with 192 GB at $1.85 an hour on demand, far more card than it needs.
The license field is empty. Open access covers fetching the weights, nothing more: there is no stated grant for commercial use, modification or redistribution, so a buyer needs written terms from the publisher and a read of the base model's terms. The one reported number is a 23.18 test word error rate on the publisher's own dataset, so measure it on your own Hebrew audio.
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the private datasets in 2 stages - firstly was fine-tuned on a small dataset with good samples Then the obtained model was fine-tuned on a large dataset with the small good dataset, with various samples from different sources, and with an unlabeled dataset that was weakly labeled using a previously trained model. (weakly labeled data wasn't used in validation set) on small dataset on large dataset on small dataset on large dataset The following hyperparameters were used during training: - learningrate: 0.0003 - trainbatchsize: 8 - evalbatchsize: 8 - distributedtype: multi-GPU - numdevices: 2 - gradientaccumulationsteps: 4…
Excerpt from the card by Vladimir Gurevich.
19 files, 2.5 GB in total. The weights are 3 files totalling 2.5 GB in bin, safetensors.
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 1.3 GB | 75c4e535a625 |
| pytorch_model.bin | Weights | 1.3 GB | 5f4ee7fbf1fd |
| training_args.bin | Weights | 3.1 KB | dfeb10e40336 |
| added_tokens.json | Configuration | 23 B | — |
| all_results.json | Configuration | 403 B | — |
| config.json | Configuration | 2.0 KB | — |
| eval.py | Configuration | 4.9 KB | — |
| preprocessor_config.json | Configuration | 214 B | — |
| run_train.py | Configuration | 36.4 KB | — |
| special_tokens_map.json | Configuration | 1.3 KB | — |
| train_results.json | Configuration | 197 B | — |
| trainer_state.json | Configuration | 26.4 KB | — |
| validation_results.json | Configuration | 227 B | — |
| README.md | Documentation | 10.4 KB | — |
| run_train.sh | Other | 1.2 KB | — |
| .gitattributes | Repository | 1.2 KB | — |
| .gitignore | Repository | 13 B | — |
| tokenizer_config.json | Tokenizer | 288 B | — |
| vocab.json | Tokenizer | 295 B | — |
Released by Vladimir Gurevich through its official repository on Hugging Face.
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 |
|---|---|---|---|---|---|
| Custom Dataset | Task Automatic Speech RecognitionMetric Test WERComparison conditions not established | 23.18 | imvladikon Publisher reported |
Evaluated revision not stated | — |
| Precision | Weights in memory |
|---|---|
| As published | 2.5 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 (315M 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.
This Python package provides an efficient way to perform forced alignment between text and audio using Hugging Face's pretrained models. it also features an improved implementation to use much less memory than TorchAudio forced alignment API. The model checkpoint uploaded here is a conversion from torchaudio to HF Transformers for the MMS-300M checkpoint trained on forced alignment dataset
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the FTSpeech dataset, being a dataset of 1,800 hours of transcribed speeches from the Danish parliament. The model achieves the following WER scores (lower is better): The use of this model needs to adhere to this license from the Danish Parliament.
You can test this model online with the Space for Romanian Speech Recognition The model ranked TOP-1 on Romanian Speech Recognition during HuggingFace's Robust Speech Challenge: This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the Common Voice 8.0 - Romanian subset dataset, with extra training data from Romanian Speech Synthesis dataset. Without the 5-gram Language Model optimization, it achieves the following results on the evaluation set (Common Voice 8.0, Romanian subset, test split): The architecture is based on facebook/wav2vec2-xls-r-300m with a speech recognition CTC head and an added 5-gram language model (using pyctcdecode and kenlm) trained on the Romanian…
ATTENTION! Metrics (float16) using evaluate library with batchsize=1
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…
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the commonvoice 8.0 dataset as well as other datasets listed below. It achieves the following results on the evaluation set: The eval.py script results using a LM are: Fine-tuned facebook/wav2vec2-large-xlsr-53 on Czech using the Common Voice dataset. 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 using the attached eval.py script: The Common Voice 8.0 train and validation datasets were used for training, as well as the following datasets: - Šmídl, Luboš and Pražák, Aleš, 2013, OVM – Otázky…