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.
Open-weight model · Speech recognition
whisper-large-v3-turbo
by OpenAI openai/whisper-large-v3-turbo
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.
Parameters809M
Context—
Weights1.6 GB
Licensemit
AccessOpen weights
Monthly Downloads6.8M
Runs On
What it takes to serve whisper-large-v3-turbo (809M 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 | 1.6 GB | 1.9 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 0.8 GB | 1.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.4 GB | 0.5 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.
SAVRN's Notes on whisper-large-v3-turbo
The lineage explains the size. This speech recognition build descends from openai/whisper-large-v3, pruned and fine-tuned with the decoding layers cut from 32 to 4, and lands at 809M parameters needing 1.9 GB at 16-bit, 1.0 GB at 8-bit or 0.5 GB at 4-bit. Against the cheapest listed rental, one 192 GB MI300X at $1.85 an hour, the design question we ask is how many audio streams to stack on one card, not whether it fits.
MIT asks for almost nothing: keep the copyright and permission notices, and commercial use, modification and redistribution are all yours. The reported evaluations are third-party figures and uneven by source, a mean word error rate of 7.83 with 2.1 on LibriSpeech clean and 16.13 on AMI, so pick the row that sounds like your audio. No context length is listed, so the audio window is a question for the publisher's documentation.
Model Card
By OpenAI, published under mit, revision 41f01f3fe87f.
Whisper
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.
Disclaimer: Content for this model card has partly been written by the Hugging Face team, and partly copied and pasted from the original model card.
Usage
Configuration
- Architecture
- WhisperForConditionalGeneration
- Layers
- 32
- Vocabulary size
- 51,866
- Stored precision
- float16
- Model type
- whisper
Identity and Version
- Repository
- openai/whisper-large-v3-turbo
- Publisher
- OpenAI
- Task
- Speech recognition
- Modality
- Audio
- Library
- transformers
- Parameters
- 809M parameters
- Languages
- en, zh, de, es, ru, ko, fr, ja
- Revision
- 41f01f3fe87f28c78e2fbf8b568835947dd65ed9
- First published
- 2024-10-01
- Last updated
- 2024-10-04
Files and Weights
13 files, 1.6 GB in total. The weights are 1 file totalling 1.6 GB in safetensors.
Weights1 file · 1.6 GB
Configuration6 files · 94.9 KB
Tokenizer4 files · 4.5 MB
Documentation1 file · 21.2 KB
Repository1 file · 1.5 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 1.6 GB | 542566a422ae |
| added_tokens.json | Configuration | 34.6 KB | — |
| config.json | Configuration | 1.3 KB | — |
| generation_config.json | Configuration | 3.8 KB | — |
| normalizer.json | Configuration | 52.7 KB | — |
| preprocessor_config.json | Configuration | 340 B | — |
| special_tokens_map.json | Configuration | 2.2 KB | — |
| README.md | Documentation | 21.2 KB | — |
| .gitattributes | Repository | 1.5 KB | — |
| merges.txt | Tokenizer | 493.9 KB | — |
| tokenizer.json | Tokenizer | 2.7 MB | — |
| tokenizer_config.json | Tokenizer | 282.8 KB | — |
| vocab.json | Tokenizer | 1.0 MB | — |
License and Download
- License
- mit
- Access
- Open weights, no gate
- Download size
- 1.6 GB
Download from OpenAI
Released by OpenAI through its official repository on Hugging Face. Read the license.
Built From
- Derived from openai/whisper-large-v3
- Described by arXiv:2212.04356
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 |
|---|---|---|---|---|---|
| hf-audio/open-asr-leaderboard | Task ami_werMetric ami_werComparison conditions not established | 16.13 | open-asr-leaderboard Reported by a third party |
Evaluated revision not stated | 2024-10-01 |
| hf-audio/open-asr-leaderboard | Task earnings22_werMetric earnings22_werComparison conditions not established | 11.63 | open-asr-leaderboard Reported by a third party |
Evaluated revision not stated | 2024-10-01 |
| hf-audio/open-asr-leaderboard | Task gigaspeech_werMetric gigaspeech_werComparison conditions not established | 10.14 | open-asr-leaderboard Reported by a third party |
Evaluated revision not stated | 2024-10-01 |
| hf-audio/open-asr-leaderboard | Task librispeech_clean_werMetric librispeech_clean_werComparison conditions not established | 2.1 | open-asr-leaderboard Reported by a third party |
Evaluated revision not stated | 2024-10-01 |
| hf-audio/open-asr-leaderboard | Task librispeech_other_werMetric librispeech_other_werComparison conditions not established | 4.24 | open-asr-leaderboard Reported by a third party |
Evaluated revision not stated | 2024-10-01 |
| hf-audio/open-asr-leaderboard | Task mean_werMetric mean_werComparison conditions not established | 7.83 | open-asr-leaderboard Reported by a third party |
Evaluated revision not stated | 2024-10-01 |
| hf-audio/open-asr-leaderboard | Task rtfxMetric rtfxComparison conditions not established | 200.19 | open-asr-leaderboard Reported by a third party |
Evaluated revision not stated | 2024-10-01 |
| hf-audio/open-asr-leaderboard | Task spgispeech_werMetric spgispeech_werComparison conditions not established | 2.97 | open-asr-leaderboard Reported by a third party |
Evaluated revision not stated | 2024-10-01 |
| hf-audio/open-asr-leaderboard | Task tedlium_werMetric tedlium_werComparison conditions not established | 3.57 | open-asr-leaderboard Reported by a third party |
Evaluated revision not stated | 2024-10-01 |
| hf-audio/open-asr-leaderboard | Task voxpopuli_werMetric voxpopuli_werComparison conditions not established | 11.87 | open-asr-leaderboard Reported by a third party |
Evaluated revision not stated | 2024-10-01 |
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 1.6 GB |
| 16-bit | 1.6 GB |
| 8-bit | 0.8 GB |
| 4-bit | 0.4 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Built on This Model
- Adapter ofwhisper-large-v3-turbo-hindi
- Derived fromwhisper-large-v3-turbo-hindi
Compare whisper-large-v3-turbo
Questions About whisper-large-v3-turbo
How much GPU memory does whisper-large-v3-turbo need?
About 1.9 GB at 16-bit and 0.5 GB at 4-bit: the weights (809M parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run whisper-large-v3-turbo 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 whisper-large-v3-turbo commercially?
Yes. whisper-large-v3-turbo is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.
Similar Models
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…
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: This is one of several Wav2Vec-models our team created during the hosted Robust Speech Event. This is the complete list of our models and their final scores: In parallel with the event, the team also converted the Norwegian Parliamentary Speech Corpus (NPSC) to the NbAiLab/NPSC in Dataset format and used that as the main source for training. We have released all the code developed during the event so that the Norwegian NLP community can build upon it when…
/ Improve list spacing / / Badge alignment consistency / Nemotron 3.5 ASR is a multilingual, streaming Automatic Speech Recognition (ASR) model engineered to deliver high-quality multilingual transcription across both low-latency streaming and high-throughput batch workloads. Developed by NVIDIA, this 600M parameter model transcribes speech into text with native support for punctuation and capitalization, and offers runtime flexibility with configurable chunk sizes, including 80ms, 160ms, 320ms, 560ms, and 1120ms. By leveraging a state-of-the-art Cache-Aware FastConformer-RNNT architecture, the model eliminates redundant overlapping computations common in traditional "buffered" streaming.…