yeagersthlm/piper-voice-sv-alma

Alma — Swedish Piper TTS Voice (medium)

Retrained Swedish text-to-speech voice for Piper.

Maintainer: Daniel Nylander (@yeagersthlm)

Quality: 8/10

Trained for 10,000 epochs on 1,875 audio samples using an NVIDIA RTX 4080.

Listen

🎧 1-minute demo (MP3)

Optimized Inference Parameters

Usage

Standalone Piper

echo 'Hej, hur mår du idag?' | piper \
  --model sv_SE-alma-medium.onnx \
  --config sv_SE-alma-medium.onnx.json \
  --output_file output.wav \
  --noise-scale 0.8 --noise-w 0.6 --length-scale 0.9

With GNOME Spiel & Orca (accessibility)

Alma integrates with libspiel — GNOME's modern speech framework — via speech-provider-piper.

Architecture:

Orca (screen reader)
  └─ libspiel (GObject Introspection client library)
       └─ D-Bus → ai.piper.Speech.Provider (speech-provider-piper)
            └─ Piper VITS engine + ONNX model
                 └─ Voice files in /usr/share/piper-voices/

libspiel replaces the aging speech-dispatcher with a modern D-Bus-based approach:

• Each voice is an independent D-Bus service — auto-activated on demand
• No central daemon needed
• Hot-swappable voices without restarting
• GStreamer pipeline for audio output
• Full GObject Introspection bindings (Python, JS, Vala, etc.)

Install everything from our PPA:

# Add repository
wget -qO- https://yeager.github.io/debian-repo/yeager-repo-key.asc | \
  sudo tee /usr/share/keyrings/yeager-repo.asc > /dev/null
sudo tee /etc/apt/sources.list.d/yeager.sources << SOURCES
Types: deb
URIs: https://yeager.github.io/debian-repo/
Suites: stable
Components: main
Signed-By: /usr/share/keyrings/yeager-repo.asc
SOURCES

sudo apt update

# Install speech stack + Alma voice
sudo apt install libspiel-1.0-0 speech-provider-piper piper-voice-sv-alma

# Optional: development headers and GObject Introspection
sudo apt install libspiel-dev gir1.2-spiel-1.0

Test with spiel CLI:

spiel speak "Hej! Det här är Alma, en svensk röst."

Available packages:

Download

• DEB package: sudo apt install piper-voice-sv-alma (PPA)
• GitHub Release: v1.0.0-alma
• Direct download: sv_SE-alma-medium.onnx (61 MB)

Training Details

• Base model: sv_SE-alma-medium (Piper VITS)
• Dataset: NST Swedish TTS Dataset (CC BY 4.0)
• Samples: 1,875 audio clips
• Epochs: 10,000 (checkpoints every 100)
• GPU: NVIDIA RTX 4080
• Precision: 32-bit
• Real-time factor: ~0.025x (40× faster than real-time)

Links

• Blog post: Alma, Spiel och Orca — Swedish accessibility guide
• speech-provider-piper — Piper D-Bus provider
• libspiel — GNOME speech framework
• Piper — Neural TTS engine

License

CC BY 4.0 (same as training data)