Truly Local Piper Multilang TTS (secure)

Local offline text-to-speech via Piper TTS. Self-contained setup, automatic language detection, per-call voice selection. Extensible to any language. Writes...

local-piper-tts-multilang-secure

Description

Local (offline) text-to-speech via Piper.

Purpose: generate audio files (OGG/Opus by default) from text, fully offline. No sending is performed by the skill — sending is handled by the agent after the file is ready.

Features

• Fully offline (no API keys)
• Self-contained setup via setup() — installs Piper into an isolated venv, no system-wide changes
• Automatic language detection for 20+ languages with English as default
• Per-call voice selection via voice parameter
• On-demand voice download via downloadVoices() — no models bundled, choose what you need
• Voice removal via removeVoice() — clean up voices you no longer want
• Extensible: add any language by installing a Piper .onnx model
• Writes outputs into OpenClaw workspace

First-run flow — full agent procedure

Follow this sequence exactly when the user asks to use TTS for the first time in a setup context.

Step 1 — check status

const s = await status();

Step 2 — install Piper if needed

If s.stage is not-setup or no-piper:

• Tell the user: "To use local TTS I need to install piper-tts into the skill's venv (~30 seconds, one-time). OK to proceed?"
• Wait for confirmation, then call setup().
• If setup returns a step containing "WARNING: espeak-ng not found", relay the warning and install instructions to the user.
• Call status() again after setup completes.

Step 3 — offer voice download if no models present

If s.stage is no-model (Piper installed but no .onnx files):

3a. Offer English defaults: Explain that two English voices are available as defaults (~65 MB each):

• en_US-ryan-medium — male, American
• en_US-amy-medium — female, American

Ask which they want, or both: "Which English voice(s) should I download? Ryan (male), Amy (female), or both?"

3b. Ask about other languages: After the English choice, ask: "Do you need any other languages? For example German, French, Spanish, Polish, Italian, Portuguese, Russian… Just tell me and I'll check what's available."

If the user names a language, look up the available models at https://github.com/rhasspy/piper/blob/master/VOICES.md and list the options. Download whatever the user picks using the same downloadVoices() call.

3c. Download everything at once:

const result = await downloadVoices(['en_US-ryan-medium', 'en_US-amy-medium', /* + any others */]);
// result.downloaded — succeeded
// result.failed     — [{stem, error}] if any failed

Each voice requires internet access. Download takes ~1–2 min per voice on a typical connection.

If any downloads fail:

• Check internet connectivity
• Verify the stem exists at https://github.com/rhasspy/piper/blob/master/VOICES.md
• Offer to retry

Step 4 — play samples so the user can choose

After downloading, generate a short audio sample for each downloaded voice and send it to the user.

For each voice, use a greeting in the voice's language:

• English: "Hello, I'm [name]. How can I help you today?"
• German: "Hallo, ich heiße [Name]. Wie kann ich Ihnen helfen?"
• French: "Bonjour, je m'appelle [prénom]. Comment puis-je vous aider?"
• Spanish: "Hola, me llamo [nombre]. ¿Cómo puedo ayudarte?"
• Polish: "Cześć, mam na imię [imię]. Jak mogę Ci pomóc?"
• Italian: "Ciao, mi chiamo [nome]. Come posso aiutarti?"
• Portuguese: "Olá, meu nome é [nome]. Como posso ajudar?"
• Russian: "Привет, меня зовут [имя]. Чем могу помочь?"
• For other languages: use an equivalent native greeting.

Replace [name] with the voice name (e.g. Ryan, Amy, Thorsten).

const sample = await tts({ text: 'Hello, I\'m Ryan. How can I help you today?', voice: 'en_US-ryan-medium' });
// send sample.path to the user as a voice message

Send all samples, then ask: "Which voice do you prefer? Or shall I download a different one?"

Step 5 — choose speech speed

After the user picks a voice, ask: "How fast should I speak? Normal is 100%. Some options: 125% (faster), 115% (slightly faster), 100% (normal), 80% (slower) — or tell me a percentage."

Always present speed as a percentage to the user. Never mention lengthScale directly.

lengthScale is the internal duration multiplier — lower = faster. To convert: lengthScale = 1 / (speed% / 100). Examples:

• 125% speed → lengthScale 0.8
• 115% speed → lengthScale 0.87
• 100% speed → lengthScale 1.0 (default)
• 80% speed → lengthScale 1.25

Generate a short sample at the chosen speed so the user can hear the difference:

const sample = await tts({ text: 'This is how I sound at this speed.', voice: 'chosen-voice', lengthScale: 0.8 });
// send sample.path to the user

Confirm with the user, then offer to save it permanently: "Should I save this as your default speed? It'll be used automatically every session."

If the user agrees:

await saveConfig({ lengthScale: 0.8 });

Once saved, tts() reads it from config.json in the skill directory automatically — no need to pass lengthScale on every call.

Step 6 — note the preferred voice and speed

Once confirmed, remember both voice and lengthScale for the session. Pass them to every subsequent tts() call unless the user asks to change them.

---

Before first use — always call status()

Always call status() before the first tts() call in a session to determine what is needed.

stage	Meaning	What to do
ready	Fully installed, at least one voice model present	Proceed with tts()
not-setup	Piper not installed	Ask user for confirmation, then call setup()
no-piper	Venv exists but piper binary missing	Ask user for confirmation, then call setup()
no-model	Piper installed but no voice model downloaded	Follow Steps 3–5 of first-run flow above

IMPORTANT: Always ask the user for confirmation before calling setup(). It installs the piper-tts package from PyPI into a venv inside the skill directory.

Usage

• Input: text, optional format ("ogg" or "wav"), optional voice (model stem), optional lengthScale (speech speed, default 1.0)
• Output: path to generated file (usually .ogg)

Controlling voice and language

To list installed voices, call listVoices() — returns stems of all installed .onnx models. Never assume a fixed list; it varies per user and installation.

Auto-detection (no voice param): The script detects language from the text using character and script analysis:

• Non-Latin scripts: Cyrillic (Russian, Ukrainian, Bulgarian), Greek, Arabic, Persian, Chinese, Japanese, Korean, Georgian
• Latin-script languages: Vietnamese, Polish, Romanian, Turkish, Czech, Slovak, Hungarian, Portuguese, Spanish, Catalan, German, Finnish, Scandinavian (Swedish, Norwegian, Danish), French, Italian
• Fallback: English keywords → first English model → any installed model

Auto-detection is best-effort. For reliable results with a specific language, always pass the voice parameter explicitly.

Explicit override: set PIPER_VOICE_MODEL env var to a full .onnx path (overrides everything).

When the user requests a specific voice or language:

1. Call listVoices() to see what is installed
2. Pass the matching stem as voice to tts(), e.g. voice: "en_US-amy-medium"
3. If the requested voice is not installed, offer to download it with downloadVoices([stem])

To switch back to auto-detect, omit the voice parameter.

Downloading additional voices

The user may say things like "I don't like this voice, use a female one" or "Download a German voice". When this happens:

1. Find the model at https://github.com/rhasspy/piper/blob/master/VOICES.md
2. Confirm the stem (e.g. de_DE-thorsten-medium) and call downloadVoices([stem])
3. Generate a sample and send it to the user
4. Confirm with listVoices() — the new voice is immediately usable

Removing voices

The user may say "remove that voice" or "I don't need the German voice anymore". When this happens:

1. Call listVoices() to confirm which voices are installed
2. Confirm with the user which voice to remove
3. Call removeVoice(stem) — e.g. removeVoice('de_DE-thorsten-medium')
4. Returns { removed, filesDeleted } on success
5. If the removed voice was the user's preferred voice, ask them to pick a new one

Never remove the last remaining voice without warning the user that TTS will stop working.

Changing speech speed

The user may say things like "speak faster", "too slow", or "speed it up". When this happens:

1. Ask what speed they want in %, or suggest: 125% (faster), 115%, 100% (normal), 80% (slower)
2. Convert their % to lengthScale: lengthScale = 1 / (speed% / 100)
3. Generate a short sample: await tts({ text: '...', voice: 'current-voice', lengthScale: 0.8 })
4. Send the sample and confirm
5. Offer to persist: "Save this as default?" — if yes, call saveConfig({ lengthScale: 0.8 })
6. Use the new lengthScale for all subsequent tts() calls in the session

Where files are written

• OPENCLAW_WORKSPACE/tts/ if OPENCLAW_WORKSPACE env var is set
• otherwise: ~/.openclaw/workspace/tts/

Dependencies

• python3 (3.8+) — required for setup() to create the venv
• ffmpeg — for WAV → OGG/Opus conversion
• espeak-ng — system library used by Piper internally; setup() checks for it and warns if missing. Install: sudo apt install espeak-ng (Debian/Ubuntu), sudo dnf install espeak-ng (Fedora), brew install espeak (macOS)
• At least one Piper .onnx + .onnx.json voice model pair in the skill directory

Platform support

• Linux x86_64: fully supported
• macOS x86_64 / arm64: fully supported
• Linux ARM: may require building piper-tts from source
• Windows: not supported

Remove

rm -rf ~/.openclaw/skills/local-piper-tts-multilang-secure

This removes everything: skill code, venv, and all voice models.