Voice Model Training Project

A base project for training a custom TTS voice model using your own voice recordings.

Goal

Train a Piper-compatible TTS voice model using Andrew's voice, suitable for use with the OpenClaw voice assistant.

---

Project Structure

voice-model/
├── data/
│   ├── raw/        # Original recordings (untouched)
│   ├── wavs/       # Processed 22050Hz mono WAV files (training-ready)
│   └── clips/      # Short segmented clips (if splitting long recordings)
├── scripts/        # Preprocessing, segmentation, training scripts
├── output/         # Trained model files
├── notes/          # Transcripts, session notes, todo lists
├── metadata.csv    # LJSpeech-format transcript index
└── README.md       # This file

---

Setup & Known Fixes

Before anything works, run the setup script:

bash scripts/setup.sh

This applies all required patches documented below. Safe to re-run.

Fix 1 — Telegram IPv6 / media download failure

Symptom: Telegram shows "failed to download media" when sending audio files. OpenClaw logs show:

MediaFetchError: Failed to fetch media from https://api.telegram.org/file/...: TypeError: fetch failed

Root cause: The Pi's DNS resolves api.telegram.org to an IPv6 address (2001:67c:4e8:f004::9), but IPv6 is not configured on this network. Node.js's built-in fetch (undici) uses the system resolver and times out. curl works because it has its own IPv4 fallback.

Fix: Pin api.telegram.org to its IPv4 address in /etc/hosts:

149.154.166.110 api.telegram.org

Also set in OpenClaw config (~/.openclaw/openclaw.json) under channels.telegram:

"network": {
  "autoSelectFamily": false,
  "dnsResultOrder": "ipv4first"
}

Note: The autoSelectFamily config only fixes the grammY polling connection. The /etc/hosts entry is required to fix the media download fetch path.

Verify:

getent ahosts api.telegram.org   # should show IPv4 only
node -e "fetch('https://api.telegram.org').then(r=>console.log('OK',r.status)).catch(e=>console.log('FAIL',e.cause?.code))"

Fix 2 — OpenClaw media inbound path

Symptom: Voice training capture hook fails to find audio files after download.

Root cause: OpenClaw saves inbound media to ~/.openclaw/media/inbound/, not /tmp/openclaw-media/ as initially assumed from the source code.

Fix: Hook handler.ts updated to use correct path:

const MEDIA_CACHE = join(homedir(), ".openclaw", "media", "inbound");

No script needed — already fixed in the hook source.

---

Training Format: LJSpeech

Piper (and most open TTS engines) use the LJSpeech format:

• Audio: 22050 Hz, mono, 16-bit WAV files in data/wavs/
• Metadata: metadata.csv in the format:

  filename|transcription|normalized_transcription
  

  Example:

  wavs/line_001|Hello, my name is Andrew.|Hello, my name is Andrew.
  

---

Recommended Approach

Phase 1 – Record

• Record yourself reading 500–2000 sentences (more = better quality)
• Good sources: common phrases, news articles, book excerpts
• Target: clean room, consistent mic position, 16kHz+ sample rate
• Avoid: background noise, mouth sounds, long pauses

Phase 2 – Process

• Convert to 22050Hz mono WAV (script provided in scripts/)
• Segment long recordings into short clips (2–10 seconds each)
• Transcribe each clip accurately

Phase 3 – Train

• Use Piper Training (cloud recommended for speed) or local on a GPU machine
• Piper training repo: https://github.com/rhasspy/piper-train

Phase 4 – Deploy

• Drop the trained .onnx model into /opt/voice-assistant/voices/
• Update voice_assistant.py to reference the new model

---

Tools

Tool	Purpose
ffmpeg	Audio conversion and preprocessing
whisper.cpp	Auto-transcription of recordings
piper-train	Model training
piper	Inference / testing

---

Resources

• Piper TTS: https://github.com/rhasspy/piper
• Piper Training: https://github.com/rhasspy/piper-train
• LJSpeech format: https://keithito.com/LJ-Speech-Dataset/
• Coqui XTTS (alternative, higher quality): https://github.com/coqui-ai/TTS

---

Status

• Project scaffolded
• Recording script prepared
• Sample sentences collected
• Recordings captured
• Audio preprocessed
• Transcriptions completed
• Training run
• Model deployed to voice assistant