whisper-swedish-media

whisper-swedish-media is a fine-tuned version of OpenAI's whisper-small, optimized for transcribing spoken Swedish from YouTube, TV programs, podcasts, and general media.

...

whisper-swedish-media

whisper-swedish-media is a fine-tuned version of OpenAI's whisper-small, optimized for transcribing spoken Swedish from YouTube, TV programs, podcasts, and general media.

---

🔍 Model Overview

• Base: openai/whisper-small
• Language: Swedish (sv)
• Domain: Broadcast media (TV, YouTube, podcasts)
• Training Data: ~70 hours of manually transcribed audio using long-form conventions
• Special Tags: <overlap>, <lang:English>, #eh, ((word)), etc.
• Sampling Rate: 16 kHz

---

📊 Evaluation

On a holdout set of 300 media segments:

Model Checkpoint	WER
Base (8kHz model)	1.234
Epoch 1 (Media)	0.376
Epoch 2 (Media)	0.543
Epoch 3 (Media) ✅ Best	0.205

---

🔐 Access & Licensing

This model is gated. For commercial use or full access:

• 📧 Contact: [email protected]
• 🔗 AWS Marketplace (WMRNORDIC)

To request access, click “Access repository” above and include a short message about your intended use. Access is reviewed manually within 24–48h.

---

📌 Intended Use

This model is built for:

• Media transcription (TV, YouTube, news)
• Swedish NLP benchmarking
• Broadcast monitoring and compliance tools

---

🛒 Commercial Access

This model is available via AWS Marketplace

Subscribe to unlock GPU-accelerated transcription.

---

🚀 Usage Example

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import soundfile as sf

processor = WhisperProcessor.from_pretrained("WMRNORDIC/whisper-swedish-media")
model = WhisperForConditionalGeneration.from_pretrained("WMRNORDIC/whisper-swedish-media")

audio, rate = sf.read("your_file.wav")
inputs = processor(audio, sampling_rate=rate, return_tensors="pt")
generated_ids = model.generate(inputs.input_features)
transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

print("Transcription:", transcription)