↓Skip to main content

STT Options For OpenVoiceOS

·5 mins
Archived (2024). Some steps here are out of date. For current options, see STT Options for OpenVoiceOS in 2026.

OpenVoiceOS, or OVOS, is the spiritual successor (and fork) to the Mycroft voice assistant. It is a privacy-focused, open-source voice assistant that you can run on your own hardware. OVOS has a plugin system that allows you to swap out the default speech-to-text (STT) engine for one that you prefer.

The plugin system, while powerful, can be confusing due to the sheer number of options available. This post will cover some of the STT options available to you and when you might use them.

Running Plugins Directly #

While the open source community typically runs voice assistants on Raspberry Pi or other low-cost hardware, some run their voice assistants on more powerful hardware. If you have a powerful machine, or the model doesn’t take much to run, you can run your STT plugin directly on your voice assistant hardware.

You may notice that there aren’t any recommendations for STT plugins that run on Raspberry Pi. While you could run citrinet, VOSK, or smaller fasterwhisper models on a Raspberry Pi, the performance will be too slow to use comfortably in a voice assistant. If you have strong privacy needs, hosting a local STT server on a more powerful machine and pointing your voice assistant to it is a better option. Otherwise, you might consider using a cloud-based STT service such as Microsoft Azure STT.

Here are the STT plugins directly maintained by OVOS as of this writing (late November 2024):

Hardware/EnvironmentModelGitHub LinkNotes
GPU-OptimizedNeMo with GPUovos-stt-plugin-nemoBest performance with GPU acceleration, but many models exist optimized for CPU
GPU-Optimizedfasterwhisperovos-stt-plugin-fasterwhisperFast and accurate with GPU acceleration, multilingual models available
GPU-OptimizedMeta MMSovos-stt-plugin-mmsGPU recommended for optimal performance
GPU-Optimizedwav2vec2ovos-stt-plugin-wav2vec2Benefits from GPU acceleration, but not required
CPU-CapableVOSKovos-stt-plugin-voskWorks well on CPU, suitable for smaller devices, not accurate and not recommended for most use cases .
CPU-Capablecitrinetovos-stt-plugin-citrinetONNX-converted NeMo model for CPU usage. Only citrinet can be exported to ONNX, and this plugin cannot run other NeMo models.
API-Based/CloudChromiumovos-stt-plugin-chromiumUses deprecated API, very fast but unsupported by Google
API-Based/CloudMicrosoft Azure STTovos-stt-plugin-azureCloud-based service, usage-based cost, fast and high quality but not private
Language-SpecificprojectAINA-remoteovos-stt-plugin-projectAINA-remoteCatalan models
Language-SpecificHiTZovos-stt-plugin-HiTZBasque models
Language-SpecificNòsovos-stt-plugin-nosGalician models
Language-SpecificIberian peninsula fasterwhisperovos-stt-plugin-fasterwhisper-zuazoOptimized for Iberian languages

OVOS offers many STT plugins as proofs-of-concept in their OVOS Hatchery Git organization. The plugins in this repository worked as of their last commit, but are not guaranteed to work in perpetuity or with the latest changes to OVOS. If you didn’t see an STT engine you wanted to try then you might see it in the Hatchery, ready to be maintained by someone with time and interest.

OVOS-STT-HTTP-Server #

If you want to run multiple STT engines at once, or if you want to run your STT engine on a different machine from your voice assistant, you can use the OVOS STT HTTP Server. This server puts a consistent API on the front of any supported OVOS STT plugin, enabling you to run your personal, private speech-to-text on a different machine from your Voice Assistant. It also means you are not constrained by the hardware on your Voice Assistant - you could run FasterWhisper on a machine with a GPU, for example, and use it as your primary STT engine. Finally, you can use a single STT server for a number of local assistants, allowing you to take advantage of a more powerful machine for all your STT needs.

In addition to usage with voice assistants, the REST API can be used for any custom application that requires speech-to-text. For example, you could use it to automatically transcribe podcasts overnight when no one would typically be using the server. You could also use it to transcribe notes you took that day from meetings or lectures.

STT Plugins From Neon #

Neon.AI is a downstream partner of OVOS and a former Mycroft channel partner. They maintain a number of STT plugins that are compatible with OVOS, though some are for specialized use cases. Here are some of the plugins available from Neon:

Hardware/EnvironmentModelGitHub LinkNotes
API-Based/CloudCustom NeMo citrinetneon-stt-plugin-nemo-remotePoint at a remote server with a custom trained NeMo model which can run on a Raspberry Pi 4 or higher, though it’s a bit slow
API-Based/CloudGoogle Cloud STTneon-stt-plugin-google_cloud_streamingGoogle Cloud STT models through their API, not private
CPU-OptimizedCustom NeMo citrinetneon-stt-plugin-nemoCustom trained NeMo model which can run on a Raspberry Pi 4 or higher, though it’s a bit slow
API-Based/CPU-OptimizedCustom NeMo citrinetneon-stt-nemoCustom trained NeMo model which can run on a Raspberry Pi 4 or higher, though it’s a bit slow. This plugin has streaming capabilities to work on low-powered hardware (think ESP32 or Raspberry Pi Zero 2W)
CPU-OptimizedSileroneon-stt-plugin-sileroSilero models, optimized for CPU usage, can run on a Raspberry Pi 4 or higher but not very accurate. Not recommended for anything but POCs.

There are a number of other Neon STT plugins, but they are either archived or use models that are no longer maintained or supported.

Conclusion #

Questions? Comments? Feedback? Let me know on the Open Conversational AI Forums or OVOS support chat on Matrix. I’m available to help and so is the rest of the community.