Speaches

Speaches is an open-source, OpenAI API-compatible server for audio AI workloads. Described by its maintainers as "Ollama, but for TTS/STT models," it gives teams a fully self-hosted alternative to the OpenAI Audio API with no per-minute fees, no vendor lock-in, and no data leaving their own infrastructure. With over 3,300 GitHub stars and active development, Speaches is a production-ready choice for privacy-conscious deployments.

Common Use Cases

Speaches fits naturally into a wide range of workflows. Development teams use it as a local OpenAI Audio API replacement during testing, eliminating API costs and network round-trips in CI pipelines. Customer-facing applications embed it to power voice interfaces, call transcription, or automated accessibility features — all without sending audio to third-party services. Content creators and podcasters run batch transcription jobs against their own media libraries. Enterprises with strict data-residency requirements deploy Speaches to keep all audio processing within a controlled environment. Researchers fine-tune or evaluate different whisper checkpoints by swapping the model via the REST API, with no redeployment needed.

Key Features

• OpenAI Audio API compatibility: Implements /v1/audio/transcriptions, /v1/audio/speech, and /v1/realtime — any SDK or tool already integrated with OpenAI works without code changes
• faster-whisper transcription: High-accuracy, multilingual speech recognition with streaming output via Server-Sent Events for low-latency applications
• Kokoro and Piper TTS: Natural-sounding text-to-speech using models that run entirely on-device; Kokoro-82M is ranked #1 in the TTS Arena
• Dynamic model management: Models load on first request and unload after a configurable idle TTL, keeping RAM usage lean; preloading is available for latency-sensitive deployments
• Gradio web UI: Built-in browser interface for testing transcription and synthesis without any additional tooling
• API key authentication: Single key protects all endpoints while leaving /docs and the OpenAPI schema publicly accessible
• Realtime API support: WebSocket-based realtime audio interaction compatible with the OpenAI Realtime API spec
• HuggingFace model registry: Any faster-whisper or ONNX TTS model on HuggingFace can be loaded by model ID — no image rebuilds required

Why deploy Speaches on Hostinger VPS

Running Speaches on a Hostinger VPS puts your audio processing infrastructure entirely under your control. Audio data — which can include sensitive conversations, medical dictation, or confidential business calls — never touches a third-party API. Hostinger VPS plans offer predictable monthly costs with no per-minute transcription charges, which can add up quickly at scale. The HuggingFace model cache is persisted in a named Docker volume, so models survive restarts and upgrades without re-downloading gigabytes of weights. You can scale vertically by upgrading your plan as usage grows, and because Speaches exposes a standard REST API, it plugs into existing infrastructure — dashboards, monitoring stacks, and downstream services — with minimal configuration. Deploying through Hostinger's one-click Docker template takes seconds: the container starts, the Gradio UI becomes available immediately, and your first transcription or synthesis request automatically fetches whichever model you need directly from HuggingFace, cached permanently on disk for every subsequent use.