Hviske-tiske Open-Source Danish ASR Model - Free Deployment, 4x Faster with Unchanged Accuracy

Hviske Tiske

Developed by pluttodk

Currently the fastest Danish ASR model, a distilled version of hviske-v2, achieving approximately 4x speed improvement while maintaining the same accuracy.

Speech Recognition OtherOpen Source License:Openrail #Danish speech recognition #High-speed transcription #Low VRAM consumption

Downloads 139

Release Time : 3/20/2025

Model Overview

An efficient model optimized for Danish automatic speech recognition, suitable for fast and accurate speech transcription tasks.

Model Features

Ultra-fast transcription

4x faster than the original hviske-v2, processing 30-minute audio in just 17 seconds

Hardware efficiency

50% reduction in GPU VRAM usage, lowering hardware requirements

Accuracy maintained

Significant speed improvement while maintaining comparable recognition accuracy to the original model

Long audio optimization

Specially optimized for long audio files, excelling with files over 30 minutes

Model Capabilities

Danish speech transcription

Real-time speech recognition

Long audio processing

Use Cases

Speech transcription

Automated meeting minutes

Automatically transcribe business meeting recordings

Quickly generate text records with accuracy comparable to manual transcription

Media subtitle generation

Automatically generate subtitles for Danish video content

Efficiently process long video files, significantly reducing production time

Speech analysis

Customer service call analysis

Batch process customer service call recordings

Quickly extract conversation text for quality analysis

🚀 HviskeTiske

Hviske-Tiske is arguably the fastest Danish Automatic Speech Recognition (ASR) model. It's a distilled version of hviske-v2, offering approximately 4x speed improvement over hviske-v2 with minimal accuracy loss, making it the fastest and nearly the most precise open-source Danish transcription model.

🚀 Quick Start

Hviske-Tiske is a remarkable advancement in Danish ASR. It significantly boosts the transcription speed, especially for long audio files, while being more hardware-efficient.

✨ Features

Turbo Model for CTranslate2

The model converts syvai/hviske-v2 to the CTranslate2 model format, enabling much faster transcriptions. It can be used in Python libraries that support CTranslate2, such as faster-whisper.

Performance and Efficiency

Speed: Offers about 4x speed improvement over hviske-v2 with minimal accuracy loss.
Hardware: Requires only half the VRAM on GPU compared to other models, reducing hardware requirements.
Long Audio Files: Shows a 17-second difference per file for long audio (30min +).

📦 Installation

No specific installation steps are provided in the original document.

💻 Usage Examples

Basic Usage

from faster_whisper import WhisperModel

model = WhisperModel("pluttodk/hviske-tiske")

segments, info = model.transcribe("audio.mp3")
for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))

📚 Documentation

Conversion

The original model is converted with the following logic Github. The model weights are saved in FP16 for optimal performance. Note that the model weights are saved in FP16. This type can be changed when the model is loaded using the compute_type option in CTranslate2.

Performance Evaluation

Performance is evaluated on alexandrainst/coral, where all audio clips are 5 seconds long. Simple tests suggest that the performance trends hold for larger audio files. (Note that faster-whisper-large-v3-turbo-ct2 performs best for 30-second audio files).

Performance Plot

Performance

Speed Test

The speed is tested on the following audio lengths, with CER and WER distributions similar to the above plot:

5 seconds
10 seconds
30 seconds
1 minute
10 minutes
30 minutes

Speed

📄 License

The model is released under the OpenRAIL license.

Property	Details
Model Type	Automatic Speech Recognition
Training Data	`alexandrainst/coral`
Base Model	`syvai/hviske-v2`, `openai/whisper-large-v3`
Metrics	WER, CER
Pipeline Tag	Automatic Speech Recognition

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご