speech new

speech

A text-to-speech workbench for monome norns. Enter or load a phrase, render it with eSpeak NG, flite, or Piper, audition it, and save the result as a WAV sample.

This began as an idea for a nota bene voice. eSpeak has to synthesize new audio when its text or settings change, so it cannot provide the deterministic note-on timing expected from an ordinary nb voice. This script makes that rendering step explicit and produces samples that can be used in a sampler or sample-based nb voice afterward.

Requirements

When SoX or every backend is missing, the script opens an installer screen that offers to install them (K3 to install, K2 to cancel). PARAMETERS > SPEECH > install dependencies opens it later for flite, Piper and the default Piper voice. Installing needs root: norns has passwordless sudo, and on desktop a graphical password prompt (pkexec) is used. If neither works, the screen shows the command to run yourself. The manual steps below do the same thing.

Install SoX and at least one speech backend. eSpeak NG and flite are available from the norns package repository:

sudo apt update
sudo apt install espeak-ng flite sox

For Piper on desktop norns, install the isolated Python package and put the model and its adjacent JSON configuration in a permanent location:

pipx install piper-tts
mkdir -p ~/.local/share/piper/voices
install -m 0644 ~/Downloads/en_GB-alan-low.onnx* \
  ~/.local/share/piper/voices/

This Alan voice path is the script default. Other Piper .onnx models can be selected from PARAMETERS.

Current piper-tts Python wheels do not support the 32-bit ARMv7 norns Shield. On a Shield running norns 260616, install Piper's standalone ARMv7 release instead:

mkdir -p ~/.local/share/piper/runtime
curl -fL \
  https://github.com/rhasspy/piper/releases/download/2023.11.14-2/piper_linux_armv7l.tar.gz \
  -o /tmp/piper_linux_armv7l.tar.gz
tar xzf /tmp/piper_linux_armv7l.tar.gz \
  -C ~/.local/share/piper/runtime
sudo ln -sfn ~/.local/share/piper/runtime/piper/piper \
  /usr/local/bin/piper

Copy the model files from the desktop with:

ssh we@norns.local 'mkdir -p ~/.local/share/piper/voices'
scp ~/Downloads/en_GB-alan-low.onnx* \
  we@norns.local:/home/we/.local/share/piper/voices/

Install

From maiden:

;install https://github.com/vehka/speech

Then refresh the script list and launch speech.

Controls

• E1 selects phrase 1-9. The selected phrase number appears in the header.
• E2 selects a setting for the active speech backend.
• E3 changes the selected setting.
• K2 renders the current phrase and previews it once. Press K2 during playback to stop it.
• K3 renders the current phrase to the norns audio library.
• K1 > PARAMETERS selects the backend, loads text, and edits phrases, voice, output name, and synthesis settings.

Preview and save are also available as triggers in the parameters menu.

Parameters

• phrase selects the active phrase slot. Phrase 1 defaults to I am Norns.; phrases 2-9 default to empty.
• PHRASES opens a submenu containing all nine phrases. Select a numbered phrase with K3 to open the wrapped text editor.
• load text file opens the norns file selector and replaces the active phrase with the selected text file's contents.
• eSpeak voice lists the voices reported by espeak-ng --voices.
• flite voice lists the voices reported by flite -lv.
• Piper model selects an ONNX voice model. Its .onnx.json configuration must be in the same directory.
• output name is sanitized to a filename; .wav is optional.

The phrases and output name use the same text editor.

Text Editor

• With a keyboard connected, the editor shows five wrapped lines. Type to insert at the cursor; use Left/Right, Home/End, Backspace, and Delete to edit. Enter saves and Escape cancels.
• Without a keyboard, the editor shows three wrapped lines above the standard character picker. E2 selects a character or DEL/OK, E3 changes rows, K3 acts, and K2 cancels.
• The visible lines follow the cursor, so the insertion point and latest words remain on screen. Moving an encoder switches to the character-picker view.

Backends

• eSpeak NG provides speed, pitch, pitch range, amplitude, and word-gap controls. A pitch range of 0 removes the intonation; 50 is the normal amount. monotone speaks the whole phrase on the note chosen with monotone note (E1 to E5), for pitch-correcting or playing the sample chromatically afterward. It replaces the pitch and pitch range settings, turns off eSpeak's pitch flutter, and works with the English (en...) voices only.
• flite provides duration-stretch and pitch-shift controls. Stretch values above 1.00x speak more slowly, while lower values speak more quickly. Pitch is adjustable by +/-12 semitones.
• Piper provides length, noise, and noise-width controls. Length values above 1.00x speak more slowly. The noise controls vary voice generation; the initial values match the Alan model configuration.

Only the active backend's settings are shown in the parameters menu. You only need to install the backend you intend to use, plus SoX. eSpeak NG is the default for new installs; a pset can select any available backend explicitly.

The phrases, active phrase number, and settings are ordinary norns params, so they are stored in the script's psets. Saved audio is written separately to:

~/dust/audio/speech/

Notes

• Parameter changes affect the next K2/K3 render, not speech already playing.
• The initial eSpeak speech rate is 120 WPM.
• Monotone renders stay within a few cents of the chosen note. eSpeak's flat voice is set in whole Hz, so SoX retunes the remainder; this changes the length by under 1%.
• Output is resampled to norns' 48 kHz audio rate before previewing or saving.
• Preview files receive a short silent tail to work around norns tape playback dropping its final buffered audio. This padding is not added to saved WAVs.
• Preview uses norns' shared tape player. Starting a preview replaces any current tape playback; it does not affect tape recording.
• Rendering is normally quick for short phrases, but matron may pause briefly while the selected backend creates the WAV.
• Piper on the ARMv7 Shield loads the Alan model in about 2.5 seconds and renders at approximately real time.
• Render errors are logged to ~/dust/data/speech/speech.log.

License

MIT