Speech2Text

Speech2Text is a desktop application for transcribing and, where supported, translating speech in audio and video files. It provides one interface for local automatic speech recognition (ASR) engines and optional remote Whisper services. Local processing keeps recordings on your computer; files leave the computer only when you select the Whisper ASR Webservice or OpenAI API.

Version 2.7.0 supports a wide range of ASR models. Availability depends on the operating system and hardware:

• macOS and Windows: OpenAI Whisper, whisper.cpp, NVIDIA Parakeet through whisper.cpp, and MOSS-Transcribe-Diarize through moss-transcribe.cpp. The native whisper.cpp, Parakeet and MOSS engines are included in the macOS, Windows x64 and Windows ARM64 builds.
• Macs with Apple silicon: MLX Whisper, Parakeet MLX, Google DeepMind Gemma 3n and Gemma 4, Microsoft VibeVoice-ASR, Mistral Voxtral Mini Realtime, quantised Qwen3-ASR, Fun-ASR Nano, NVIDIA Nemotron and Sortformer, Meta Massively Multilingual Speech (MMS), and Granite Speech 4.0. These models use MLX, which requires an M1 or later.
• Other local engines: Faster Whisper is available on Macs with Apple silicon and on Windows. The official Qwen3-ASR engine is available on Apple silicon and on Windows x64 with an NVIDIA CUDA GPU. IBM Granite Speech 4.1 is available on Apple silicon and Windows, including Windows ARM64.
• All supported platforms: a remote Whisper ASR Webservice or the OpenAI API.

Features

• Transcribes common audio and video formats, including WAV, MP3, M4A, OGG, MP4, MOV and AVI.
• Provides a choice of transcript, subtitle and research-oriented output formats (output format details).
• Supports speaker diarisation (i.e. identifying who is speaking and labelling each person's contribution) with compatible models, including MOSS-Transcribe-Diarize, VibeVoice, Granite Speech 4.1 and Sortformer.
• Downloads model files on demand and shows model-download, conversion and transcription progress in a cancellable progress window.
• Shows memory guidance and effective audio-length limits for the selected engine and model.

Output formats

Speech2Text offers more than a plain transcript. The Output menu is updated for the selected engine and model, so only supported formats are shown.

• TEXT saves the transcription as an ordinary text file.
• SRT and VTT create timed subtitles for video players, presentation software and web pages.
• ASS creates styled subtitles and can show different speakers in different colours.
• JSON retains structured information supplied by the model, such as segments, timestamps, words and speaker labels.
• WTSV creates a tab-delimited word-timing file with each word's onset, offset and duration in milliseconds. A speaker column is included when the model provides speaker information.
• Praat TextGrid creates interval tiers that can be opened directly in Praat for phonetic analysis and annotation.

Whisper.cpp also provides TSV, CSV, LRC and karaoke timing output. Word-level formats such as WTSV and TextGrid are available only when the selected model provides reliable word timestamps.

Platform support

MLX-based engines require a Mac with Apple silicon (M1 or later). Parakeet through whisper.cpp and MOSS through moss-transcribe.cpp run locally on both macOS and Windows, with CPU and supported GPU acceleration. Windows x64 supports CUDA and cross-vendor Vulkan acceleration; the native Windows ARM64 build supports CPU and Vulkan. Qwen-ASR and Granite Speech 4.1 are available on supported macOS and Windows configurations.

Memory requirements vary considerably by model. A computer with at least 16 GB of memory is recommended for the wider model selection, although smaller Whisper, Parakeet, Qwen3-ASR and quantised MOSS models can use less. Speech2Text uses whisper.cpp with the base model by default. Larger Whisper models can improve transcription of noisy recordings or accented speech, but require more memory and processing time (see the transcription speed comparison).

Non-WAV media are converted to 16 kHz, 16-bit WAV before transcription. Some engines and file formats require FFmpeg (installation instructions).

Download

Current version: 2.7.0 (21 August 2026)

• macOS for Apple silicon (M1 or later, DMG)
• macOS for Intel processors (DMG)
• Windows x64 (Intel or AMD 64-bit, installer)
• Windows ARM64 (native ARM64 installer)

Key changes in version 2.7.0

• Added Gemma 4, quantised Qwen3-ASR, Fun-ASR Nano, Voxtral Mini Realtime and Nemotron 3.5 ASR models.
• Added Parakeet through whisper.cpp and MOSS-Transcribe-Diarize through moss-transcribe.cpp on macOS and Windows, including Metal, CUDA or Vulkan acceleration where supported.
• Added a native Windows ARM64 installer; Windows x64 now supports cross-vendor Vulkan and an optional CUDA backend downloaded on demand.
• Improved the responsive interface, engine/model status, progress and cancellation feedback, model memory guidance, and engine-aware task and output choices.
• Moved large optional runtimes outside the app bundle and improved model download and maintenance cleanup on macOS and Windows.

Installation and starting Speech2Text

On macOS, open the DMG and drag Speech2Text to the Applications folder. On Windows, download and run the installer that matches your computer: Speech2Text_Setup_x64.exe for a conventional Intel or AMD PC, or Speech2Text_Setup_ARM64.exe for a Windows on Arm device.

The application is not code signed, so macOS or Windows may display a security warning the first time you open it. Follow the platform-specific instructions below only after confirming that you downloaded the installer from this page.

macOS Sequoia and later

The first time you start the application on macOS (double click on the application icon in Finder), you will see the dialog box below. Click on the button "Done".

To open the app, go to System Settings, select Privacy & Security, and click Open Anyway beside the Speech2Text message.

When the confirmation dialog appears, click Open Anyway again.

Finally, authenticate when prompted. Speech2Text can then be opened normally in the future.

macOS Sonoma and earlier

The first time you start the application on macOS (double click on the application icon in Finder), you will see the dialog box below:

Please note that starting the application the first time might take a while.

To fix the security warning, right-click on the application icon in the Finder, and then select in the popup menu "Open". Next, you will see this message:

Select "Open" and then it will take a few seconds before the app starts. The security warning will now not appear anymore for this app and the application should start up quickly.

Windows

The following message will appear when you start the installer.

Click on "More info", and the following message will appear.

Click on "Run anyway" to start the installer.

Transcription speed

Tables 1 and 2 provide reference timings for transcribing an 8-minute, 37-second recording of John F. Kennedy's "Ich bin ein Berliner" speech. These measurements were made with Speech2Text 2.5.7 and are retained as a hardware comparison; version 2.7.0 includes additional engines and models that were not part of this benchmark.

Install FFmpeg

Use the instructions for your operating system to install FFmpeg. Restart Speech2Text after installation so that it can detect the executable.

macOS

Open the Terminal in macOS.

If Homebrew is not installed, follow the installation instructions at brew.sh.

Then enter the following command:

%brew install ffmpeg

Copy

Windows

You can use winget to install FFmpeg on Windows.

Open the PowerShell Terminal in Windows.

Enter the following command:

>winget install ffmpeg

Copy

Alternatively, use Scoop or Chocolatey to install FFmpeg on Windows.

Requirements and notes

• FFmpeg is required by OpenAI Whisper and MLX Whisper and may be needed to convert non-WAV media for other engines.
• Automatic language detection with Whisper-based engines requires at least 30 seconds of audio. For a shorter recording, select the language manually.
• At startup, Speech2Text contacts waltervanheuven.net to check whether a newer version is available. Updates are not downloaded or installed automatically.
• Models are downloaded when first selected, so the initial transcription with a new model takes longer and requires an internet connection.
• Model files and optional engine environments can use several gigabytes of disk space. They can be reviewed and removed from the Maintenance section in Settings.

Disclaimer

Speech2Text is provided “as is”, without warranty of any kind, express or implied. To the fullest extent permitted by law, the author shall not be liable for any claim, loss, damage or other liability arising from the software or its use. Automatic transcriptions may contain errors and should be checked before being used for research, clinical, legal or other consequential purposes.

Citation

If you use Speech2Text in your research, please cite it as follows:

van Heuven, W. J. B. (2026). Speech2Text (Version 2.7.0) [Computer software]. https://waltervanheuven.net/s2t/

Source code

The source code is available at https://github.com/waltervanheuven/speech2text.

Speech2Text was originally designed and written by Walter van Heuven. In later versions, the code was improved and updated with the assistance of AI-based coding tools.

If you have any questions, contact Walter van Heuven.