aTrain: A Localized AI Solution for Offline Speech Transcription and Speaker Diarization

aTrain is an offline speech transcription tool developed by researchers at the University of Graz in Austria. Based on OpenAI Whisper (implemented via Faster-Whisper) and pyannote.audio technologies, it supports speech recognition and speaker diarization in 99 languages. Its core advantages include fully local operation to ensure data privacy (compliant with regulations like GDPR) and compatibility with mainstream qualitative analysis software such as MAXQDA and ATLAS.ti. This open-source tool aims to provide efficient, privacy-protected speech processing solutions for users who value data sovereignty.

Most mainstream speech transcription services rely on cloud processing, which carries the risk of sensitive data leakage. Especially under the strict supervision of the EU's GDPR, researchers and institutions urgently need solutions that allow them to control data sovereignty. aTrain was developed to address this need: it is created by the Center for Business Analytics and Data Science at the University of Graz and tested in collaboration with Know-Center Graz, specifically designed for privacy-sensitive scenarios.

Speech Recognition Engine: Faster-Whisper

It uses Faster-Whisper (a high-performance version of OpenAI Whisper) developed by Guillaume Klein. While maintaining high accuracy, it improves processing speed: the highest-quality model takes approximately 3 times the audio duration to process on a mid-range business laptop.

Speaker Diarization Technology: PyAnnote.Audio

Integrated with pyannote.audio for speaker detection, it automatically distinguishes content from multiple speakers—suitable for meetings, interviews, and other scenarios without manual annotation.

• Fully Offline Processing: All operations are done locally with no data uploads, ensuring privacy and compliance.
• Multilingual Support: Covers 99 languages (e.g., Chinese, English, German, etc.). For transcription quality across different languages, refer to Whisper's WER data.
• Qualitative Software Integration: Transcribed files can be directly imported into MAXQDA, ATLAS.ti, and NVivo, with timestamped audio playback support to enhance research efficiency.
• GPU Acceleration: Equipped with an NVIDIA GPU (CUDA required) can reduce processing time to 20% of the audio duration—e.g., a 22-minute recording takes only about 4.4 minutes to complete.

The project team tested using a 22-minute dialogue video from the 2023 Bank Supervision Forum of the European Central Bank. With speaker diarization enabled, transcription took only about 4.4 minutes on an entry-level gaming laptop (with NVIDIA GPU). This performance demonstrates that the tool is suitable for post-hoc batch processing and fast-turnaround scenarios (e.g., news interviews).

aTrain is suitable for:

• Academic interviews/focus groups: Quickly obtain analyzable text records
• Meeting minutes: Structured documentation with speaker differentiation
• Media production: Journalists and podcast creators converting audio to text
• Legal/medical transcription: High accuracy and privacy protection
• Multilingual processing: International teams handling non-English content

aTrain represents an important direction for voice AI tools: balancing large model quality with data control rights. It is a highly valuable open-source solution for European research institutions, organizations handling sensitive information, and users who value data sovereignty. In the future, as local computing power improves and models are optimized, offline AI tools will become more prevalent in more fields, providing intelligent and secure productivity support.