What it does OpenTTS is a Python server that wraps eight different open-source text-to-speech systems — from neural models like Larynx and Coqui-TTS to classic rule-based engines like Festival and eSpeak — and exposes them through one HTTP API. You request a voice by name (e.g., glow-speak:en-us_mary_ann or coqui-tts:en_vctk#p228), and the server returns WAV audio. It also supports a practical subset of SSML, including voice switching mid-document, breaks, and language tags.

The interesting bit The SSML implementation is where the glue code gets clever: you can switch TTS engines within the same document, jumping from a neural English voice to a robotic-but-functional Swahili voice, or mixing German, Japanese, and Arabic sentences. The server also mimics MaryTTS’s endpoint, so existing integrations can swap in OpenTTS without code changes.

Key highlights

• Docker-first deployment with per-language images (synesthesiam/opentts:en, :de, :all, etc.)
• WAV caching with optional disk mount for repeated phrases
• Voice quality tiers (high/medium/low) for Larynx/Glow-Speak, useful on Raspberry Pi
• Multi-speaker model support via #speaker_id syntax (currently Coqui-TTS English)
• Web UI and OpenAPI test page at localhost:5500

Caveats

• Build-from-source requires re-initializing QEMU/docker buildx after every reboot (documented, but annoying)
• Some Festival voices use legacy ISO-8859 encodings; Russian requires automatic transliteration to Latin
• eSpeak is included by default and explicitly described as “robotic”; use --no-espeak to exclude it

Verdict Worth a look if you’re running home automation, accessibility tools, or any project that needs offline TTS across multiple languages without managing eight separate toolchains. Skip it if you need production-grade cloud voices or deep SSML control — this is a pragmatic unification layer, not a voice quality breakthrough.