AI Transcription & Voice, Hosted in France
A speech-to-text and text-to-speech server compatible with OpenAI powered by Speaches, installed and maintained by DINAO. Your recordings and generated voices stay on our French servers — they never go to OpenAI.
What is Speaches?
Speaches is a server compatible with the OpenAI API dedicated to transcription, translation, and speech generation. Speech recognition (speech-to-text) is handled by faster-whisper, while speech synthesis (text-to-speech) relies on Piper and Kokoro. The project's goal: to be an « Ollama, but for TTS/STT models ».
Because its API is compatible with OpenAI, all tools and SDKs designed for OpenAI's audio API work with Speaches without modification. Transcription is streamed via streaming through SSE as the audio is processed : no need to wait for the file to finish to receive the text.
Self-hosted, Speaches guarantees total data privacy, eliminates per-minute costs, and allows audio processing without depending on a cloud API. Multiple models can coexist and be loaded on demand, from the most CPU-light to large GPU-accelerated models.
Host Speaches at DINAO
Resource tiers compatible with Speaches prerequisites (minimum 2 cœurs / 4 Go / 10 Go (cache des modèles Hugging Face)). Hosted in France, fully managed.
- 2 dedicated vCPU
- 4 Go RAM
- 40 GB NVMe
- Daily backups
- Managed & monitored by DINAO
- 4 dedicated vCPU
- 8 Go RAM
- 80 GB NVMe
- Daily backups
- Managed & monitored by DINAO
- 8 dedicated vCPU
- 16 Go RAM
- 160 GB NVMe
- Daily backups
- Managed & monitored by DINAO
1 hidden tier(s) (insufficient resources for this app) : Découverte
This application uses AI
The container hosts the application, not the AI engine (which requires dedicated GPUs) : inference is handled externally, with your own provider key. Prioritize a sovereign engine — Mistral AI, NumSpot, Scaleway, or OVHcloud AI Endpoints (France, GDPR) ; an international provider (OpenAI, Anthropic…) only if a specific capability requires it. Inference subscriptions are not included in hosting.
Technical details
You might be wondering…
Do my recordings go to OpenAI?
No. In local mode, your audio files are transcribed and synthesized directly on your DINAO instance via Whisper, Piper, and Kokoro. No sound or text is transmitted to a third-party provider.
Is Speaches compatible with OpenAI tools?
Yes. Speaches exposes an OpenAI-compatible API : SDKs and applications that use OpenAI's audio API work by simply pointing to your instance.
Which models are available?
For transcription, Whisper models (faster-whisper) from the lightest to the most accurate. For speech synthesis, Piper and Kokoro, with multiple voices. A GPU significantly accelerates large models.
Where is the data hosted?
On DINAO's infrastructure in France, in one of the available data centers. Your audio and text never leave the territory.
Is a GPU required?
No for light models, which run on CPU. For real-time transcription or large Whisper models, an optional GPU significantly improves speed.
