SpeakoFlow Mini: An 833 MB Offline Dictation Cleanup Model

Overview

SpeakoFlow Mini is a specialized software tool designed to clean up rough speech-to-text transcripts. It is an 833 MB language model that runs entirely on a user's computer without needing an internet connection. The tool is built on top of the Qwen3.5-0.8B model and uses an open license called Apache-2.0. Its main job is to fix specific errors that standard transcription tools often miss, such as correcting words that were changed mid-sentence or turning spoken symbols like "at" into actual email addresses. Because it works offline, it is perfect for users who care about privacy or work in places with poor internet access.

Benefits

SpeakoFlow Mini offers several key advantages for anyone working with voice notes or transcripts. First, it operates completely offline, meaning no data leaves the user's device. This ensures high privacy and works even without an internet connection. Second, it is very lightweight at only 833 MB, making it easy to run on standard CPUs without needing expensive hardware. Third, it is highly accurate at its specific tasks. It successfully leaves correct text untouched 92.6% of the time, which prevents accidental changes to good content. Finally, it performs as well as much larger, paid online models on specific tasks like handling long passages, but without the cost or delay of using a cloud service.

Use Cases

This tool is ideal for writers, researchers, and professionals who transcribe audio into text. It is particularly useful for cleaning up raw transcripts that contain common dictation mistakes. For example, it can fix sentences where a speaker changes their mind, like changing "Thursday" to "Friday" mid-sentence. It also converts spoken symbols into written text, turning "sam at example dot com" into a proper email address. The tool is also great for interpreting dictation commands, such as when a user says "New paragraph" to start a new section. It works best on English text and is suitable for environments where security and offline capability are priorities.

Pricing

SpeakoFlow Mini is free to use. It is distributed under the Apache-2.0 license, which allows users to download and run it without any cost. The model is available as a GGUF file and can be used with popular open-source tools like llama.cpp, Ollama, and LM Studio. Users can run it on their own hardware without paying per-use fees or subscription costs.

Vibes

The model has received positive attention for its ability to match the performance of large, hosted AI models while offering better privacy and lower costs. In tests, it achieved a statistical tie with advanced online models on overall tasks and significantly outperformed them on longer passages of 500 words or more. Users and developers appreciate that it does not require a network connection and avoids the latency of waiting for a server response. While it is English-only and may occasionally miss a word spoken by the user, it is widely regarded as a powerful solution for local dictation cleanup that outperforms other similar offline models.

Additional Information

SpeakoFlow Mini was developed by the team behind SpeakoFlow and is hosted on Hugging Face. The model was fine-tuned from the Qwen3.5-0.8B base model. It is designed to run on CPUs with a median latency of about 2.5 seconds per transcript, though using a GPU can speed this up to 0.3 seconds. The developers emphasize that the model is trained to be restrained, meaning it only makes changes when necessary and avoids altering correct text. It is available in several quantization options to balance file size and performance, with the full 833 MB version offering the highest accuracy.