README.md

---
library_name: transformers
license: mit
tags:
- automatic-speech-recognition
- audio
- speech
- whisper
- multilingual
- streaming
- coreml
- cuda
- nvidia
- apple-silicon
- on-device
---

# TheWhisper-Large-V3

## Model Summary

**TheWhisper-Large-V3** is a fine-tuned, high-performance variant of OpenAI’s Whisper Large V3 model — optimized by **TheStage AI** for **real-time**, **low-latency**, and **low-power** speech-to-text (ASR) inference across multiple platforms, including **NVIDIA GPUs** and **Apple Silicon (CoreML)**.  

It provides **streaming transcription**, **word timestamps**, and **scalable performance** for use cases like real-time captioning, meetings, and on-device voice interfaces.


## 📊 Quality Benchmarks

TheWhisper is a fine-tuned Whisper model that can process audio chunks of any size up to 30 seconds. Unlike the original Whisper models, it doesn't require padding audio with silence to reach 30 seconds. We conducted quality benchmarking across different chunk sizes: 10, 15, 20, and 30 seconds. For quality benchmarks, we used the multilingual benchmarks [Open ASR Leaderboard](https://github.com/huggingface/open_asr_leaderboard#evaluate-a-model).

<img width="1547" height="531" alt="vanilla whisper (1)" src="https://github.com/user-attachments/assets/f0c86e58-d834-4ac7-a06b-df3a7ae3e9e9" />
<img width="1547" height="458" alt="TheStage AI Whisper (1)" src="https://github.com/user-attachments/assets/17fb45a3-b33d-4c83-b843-69b0f0aa3f65" />


### 10s chunks

| Model | Mean WER |
|-------|-----------------|
| openai/whisper-large-v3-turbo | 7.81 |
| openai/whisper-large-v3 | 7.45 |
| thewhisper-large-v3-turbo | 7.88 |
| thewhisper-large-v3 | 7.8 |


### 15s chunks

| Model | Mean WER |
|-------|-----------------|
| openai/whisper-large-v3-turbo | 7.61 |
| openai/whisper-large-v3 | 7.22 |
| thewhisper-large-v3-turbo | 7.45 |
| thewhisper-large-v3 | 7.34 |

### 20s chunks

| Model | Mean WER |
|-------|-----------------|
| openai/whisper-large-v3-turbo | 7.63 |
| openai/whisper-large-v3 | 7.29 |
| thewhisper-large-v3-turbo | 7.47 |
| thewhisper-large-v3 | 7.31 |

### 30s chunks

| Model | Mean WER |
|-------|-----------------|
| openai/whisper-large-v3-turbo | 7.61 |
| openai/whisper-large-v3 | 7.32 |
| thewhisper-large-v3-turbo | 7.45 |
| thewhisper-large-v3 | 7.28 |


## Quick start
---

### Apple Usage

```python
import torch
from thestage_speechkit.apple import ASRPipeline

model = ASRPipeline(
    model='TheStageAI/thewhisper-large-v3',
    # optimized model with ANNA
    model_size='S'
    chunk_length_s=10,
    token=hf_token
)

# inference
result = model(
    "path_to_your_audio.wav", 
    max_batch_size=32,
    return_timestamps="word"
)

print(result["text"])
```

### Apple Usage with Streaming

```python
from thestage_speechkit.apple import WhisperStreamingPipeline
from thestage_speechkit.streaming import MicStream, FileStream, StdoutStream

streaming_pipe = WhisperStreaming(
    model='TheStageAI/thewhisper-large-v3',
    # Optimized model by ANNA
    model_size='S',
    # Window length
    chunk_length_s=10,
    platform='apple'
)

# set stride in miliseconds
mic_stream = MicStream(step_size_s=0.5)
output_stream = StdoutStream()

while True:
    chunk = mic_stream.next_chunk()
    if chunk:
        approved_text, assumption = streaming_pipe(chunk)
        output_stream.rewrite(approved_text, assumption)
    else:
        break
```

### Nvidia Usage (HuggingFace Transfomers)

```python
import torch
from thestage_speechkit.nvidia import ASRPipeline

model = ASRPipeline(
    model='TheStageAI/thewhisper-large-v3',
    # allowed: 10s, 15s, 20s, 30s
    chunk_length_s=10,
    # optimized TheStage AI engines
    device='cuda',
    token=hf_token
)

# inference
result = model(
    audio="path_to_your_audio.wav", 
    max_batch_size=32,
    return_timestamps="segment"
)

print(result["text"])
```

### Nvidia Usage (TheStage AI engines)

```python
import torch
from thestage_speechkit.nvidia import ASRPipeline

model = ASRPipeline(
    model='TheStageAI/thewhisper-large-v3',
    # allowed: 10s, 15s, 20s, 30s
    chunk_length_s=10,
    # optimized TheStage AI engines
    mode='S',
    device='cuda',
    token=hf_token
)

# inference
result = model(
    "path_to_your_audio.wav", 
    max_batch_size=32,
    return_timestamps="segment"
)

print(result["text"])
```

## Model Details
---

- **Developed by:** TheStage AI  
- **Model type:** Speech-to-Text (Automatic Speech Recognition)  
- **Languages:** Multilingual (same as Whisper Large V3: ~99 languages supported)  
- **License:** MIT  
- **Finetuned from:** [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3)  
- **Frameworks:** PyTorch, CoreML  
- **Supported Platforms:**  
  - NVIDIA GPUs (CUDA 11.8+)  
  - Apple Silicon (M1–M4, macOS 15+)

### Links

- **Repository:** [https://github.com/TheStageAI/TheWhisper](https://github.com/TheStageAI/TheWhisper)  
- **Demo / Docs:** [https://app.thestage.ai](https://app.thestage.ai)  
- **Weights:** [https://huggingface.co/TheStageAI/thewhisper-large-v3](https://huggingface.co/TheStageAI/thewhisper-large-v3)

---