Piper

Piper Text-to-Speech Skill

This skill enables you to use Piper TTS to convert text files or text input into natural-sounding speech audio files.

Installation

Piper has been installed via uv with Python 3.13:

uv tool install --python 3.13 piper-tts

The piper executable is located at: /Users/katiemulliken/.local/bin/piper

Voice Models

Voice models are stored in ~/piper-voices/.

Currently installed voices:

• en_US-amy-medium: Natural-sounding US English female voice

Downloading Additional Voices

To download more voices from Hugging Face:

# List available voices at: https://huggingface.co/rhasspy/piper-voices
# Preview samples at: https://rhasspy.github.io/piper-samples/

# Download a voice (example: en_US-lessac-medium)
cd ~/piper-voices
curl -L "https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/en/en_US/lessac/medium/en_US-lessac-medium.onnx" -o en_US-lessac-medium.onnx
curl -L "https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/en/en_US/lessac/medium/en_US-lessac-medium.onnx.json" -o en_US-lessac-medium.onnx.json

Basic Usage

IMPORTANT: When working with Obsidian markdown files (.md), ALWAYS use the clean_obsidian_for_tts.py script first to remove formatting, frontmatter, and other non-speech content before converting to audio. See the "Cleaning Obsidian Files for TTS" section below.

Convert text to audio file

# From text input
echo "Hello, this is a test." | piper -m ~/piper-voices/en_US-amy-medium.onnx -f output.wav

# From a text file
piper -m ~/piper-voices/en_US-amy-medium.onnx -f output.wav < input.txt

# Using --input-file flag
piper -m ~/piper-voices/en_US-amy-medium.onnx -i input.txt -f output.wav

Play audio immediately (requires ffplay)

echo "This will play on your speakers." | piper -m ~/piper-voices/en_US-amy-medium.onnx | ffplay -

Advanced Options

Speed Control

# Default speed (1.5x faster - length-scale 0.67)
piper -m ~/piper-voices/en_US-amy-medium.onnx --length-scale 0.67 -f output.wav < input.txt

# Normal speech (1.0 is normal)
piper -m ~/piper-voices/en_US-amy-medium.onnx --length-scale 1.0 -f output.wav < input.txt

# Slower speech
piper -m ~/piper-voices/en_US-amy-medium.onnx --length-scale 1.2 -f output.wav < input.txt

Volume Control

piper -m ~/piper-voices/en_US-amy-medium.onnx --volume 1.5 -f output.wav < input.txt

Sentence Pauses

# Add 0.5 seconds of silence between sentences
piper -m ~/piper-voices/en_US-amy-medium.onnx --sentence-silence 0.5 -f output.wav < input.txt

GPU Acceleration

piper -m ~/piper-voices/en_US-amy-medium.onnx --cuda -f output.wav < input.txt

Cleaning Obsidian Files for TTS

A Python script is included to clean Obsidian markdown files for optimal text-to-speech conversion.

Script location: This skill includes clean_obsidian_for_tts.py in the same directory as this documentation.

What it removes:

• YAML frontmatter
• Markdown formatting (headers, bold, italic, strikethrough)
• Links and URLs (keeps link text)
• Obsidian wiki links [[link]]
• Images (but preserves alt-text)
• Code blocks
• HTML tags
• Emojis and special Unicode characters
• List markers
• Excessive whitespace

Enhanced Workflow: Including Image Transcriptions

For articles with images, you can create a richer audio experience by transcribing image content:

1. Download and examine images from the article using curl or web tools
2. Transcribe image content into a cleaned text file, replacing image references with detailed descriptions
3. Insert transcriptions at the image locations in your cleaned file
4. Convert to audio with piper

This ensures images are properly represented in the audio narration, making the content accessible even without visual context.

Usage:

# Clean a file and save to a new file
python3 clean_obsidian_for_tts.py input.md -o output.txt

# Clean and show statistics
python3 clean_obsidian_for_tts.py input.md -o output.txt --stats

# Clean to stdout (for piping)
python3 clean_obsidian_for_tts.py input.md

# Clean from stdin
cat input.md | python3 clean_obsidian_for_tts.py > output.txt

Complete workflow for Obsidian to audio (with default 1.5x speed):

# Step 1: Clean the markdown file
python3 clean_obsidian_for_tts.py "My Note.md" -o "My Note - Clean.txt" --stats

# Step 2: Convert to audio with piper
piper -m ~/piper-voices/en_US-amy-medium.onnx \
  -i "My Note - Clean.txt" \
  -f "My Note.wav" \
  --sentence-silence 0.3 \
  --length-scale 0.67

# Or combine in one line:
python3 clean_obsidian_for_tts.py "My Note.md" | \
  piper -m ~/piper-voices/en_US-amy-medium.onnx \
  -f "My Note.wav" \
  --sentence-silence 0.3 \
  --length-scale 0.67

Common Command Patterns

Convert a markdown file to audio

# For Obsidian/markdown files, ALWAYS clean first with the script:
python3 clean_obsidian_for_tts.py document.md | \
  piper -m ~/piper-voices/en_US-amy-medium.onnx \
  -f document.wav \
  --sentence-silence 0.3 \
  --length-scale 0.67

# Or save the cleaned version first:
python3 clean_obsidian_for_tts.py document.md -o document-clean.txt
piper -m ~/piper-voices/en_US-amy-medium.onnx -i document-clean.txt -f document.wav

Batch process multiple files

for file in *.txt; do
  piper -m ~/piper-voices/en_US-amy-medium.onnx -i "$file" -f "${file%.txt}.wav"
done

Batch convert Obsidian notes to audio

for file in *.md; do
  python3 clean_obsidian_for_tts.py "$file" | \
    piper -m ~/piper-voices/en_US-amy-medium.onnx \
    -f "${file%.md}.wav" \
    --sentence-silence 0.3 \
    --length-scale 0.67
done

Convert articles with image transcriptions to audio

For articles containing images (like screenshots, diagrams, or referenced images):

# Step 1: Download images from the article
mkdir -p /tmp/article_images
cd /tmp/article_images
curl -L "https://example.com/image1.png" -o image1.png
curl -L "https://example.com/image2.png" -o image2.png

# Step 2: View images and manually transcribe their content
# (Use your image viewer or convert images to text using OCR tools if available)

# Step 3: Create an enhanced cleaned text file
# Start with the cleaned markdown, then replace image references with detailed transcriptions
python3 clean_obsidian_for_tts.py "article.md" > cleaned_base.txt

# Edit cleaned_base.txt to insert transcriptions like:
# Image 1: [Detailed description of what appears in image1.png]
# Image 2: [Detailed description of what appears in image2.png]

# Step 4: Convert the enhanced cleaned file to audio
piper -m ~/piper-voices/en_US-amy-medium.onnx \
  -i cleaned_with_transcriptions.txt \
  -f "article_with_images.wav" \
  --sentence-silence 0.3 \
  --length-scale 0.67

Example transcription format:

Original markdown:

![Screenshot of error message](https://example.com/error.png)

In cleaned text:

Image 1: Screenshot of error message

This image shows a red error dialog box with the message "File not found error 404". The dialog contains an OK button in the bottom right. The background appears to be a Windows desktop environment.

This approach ensures all visual content is represented in the audio version, making your content fully accessible to audio listeners.

Output to a specific directory

piper -m ~/piper-voices/en_US-amy-medium.onnx -i input.txt -d ~/audio-outputs -f output.wav

Available Options

• -m, --model: Path to ONNX model file (required)
• -c, --config: Path to model config file (optional, auto-detected from .onnx.json)
• -i, --input-file: Path to input text file
• -f, --output-file: Path to output WAV file (default: stdout)
• -d, --output-dir: Directory for output files (default: current directory)
• --output-raw: Stream raw audio to stdout instead of WAV
• -s, --speaker: Speaker ID for multi-speaker models (default: 0)
• --length-scale: Speech speed multiplier (default: 1.0)
• --noise-scale: Generator noise level
• --noise-w-scale: Phoneme width noise level
• --cuda: Enable GPU acceleration
• --sentence-silence: Seconds of silence between sentences (default: 0.0)
• --volume: Volume multiplier (default: 1.0)
• --no-normalize: Disable automatic volume normalization
• --data-dir: Directory to search for voice models
• --debug: Enable debug output

Tips

1. Large files: For very large text files, consider splitting them into smaller chunks to avoid memory issues
2. Quality vs Speed: Medium quality voices offer a good balance; high quality voices are slower but more natural
3. Preprocessing: Remove special characters or formatting that might not be pronounced well
4. Performance: The CLI loads the model each time; for repeated use, consider the HTTP API server mode

Troubleshooting

Command not found

Make sure /Users/katiemulliken/.local/bin is in your PATH:

export PATH="/Users/katiemulliken/.local/bin:$PATH"

Or use the full path:

/Users/katiemulliken/.local/bin/piper [options]

Model file errors

Ensure both the .onnx model file and .onnx.json config file are in the same directory with matching names.

Resources

• Voice samples: https://rhasspy.github.io/piper-samples/
• Voice models: https://huggingface.co/rhasspy/piper-voices
• Documentation: https://github.com/OHF-Voice/piper1-gpl
• CLI docs: https://github.com/OHF-Voice/piper1-gpl/blob/main/docs/CLI.md