Free lesson

Choose a private task and separate licenses

Choose a private task and separate licenses — Local text-to-speech with Piper. Define a synthetic announcement; Inspect the engine and voice independently; Budget hardware without promising speed. Practical assignment: Lab evidence.

Demo edition. The materials are undergoing editorial review.

Topics covered

  • Define a synthetic announcement
  • Inspect the engine and voice independently
  • Budget hardware without promising speed

Define a synthetic announcement

Use a short announcement for an invented workshop, with no real attendee names, telephone numbers or recorded voices. The teaching sentence is Dutch because the inspected voice is Belgian Dutch; changing the interface language does not make that voice multilingual. Write the intended meaning beside the text: this is an exercise and the workshop begins Tuesday at ten. The task is ordinary text-to-speech with an existing voice, not learning somebody’s vocal identity. Decide where input files and output audio will live before installing anything. Choose a local directory outside cloud-synced folders and record who can access it. Local synthesis removes a cloud inference request, but package downloads, telemetry in surrounding tools, backups and shared speakers remain separate privacy paths. The expected artifact is a one-page task note naming text, language, storage and excluded personal data.

Keep the disclosure in the actual text file, not only the project README. A later listener may receive the clip without its folder, so the opening sentence must still identify the example as an exercise.

Inspect the engine and voice independently

Create a rights ledger with distinct rows for engine, voice file, direct dataset and earlier training bases. The current Piper engine license is GPL-3.0-or-later; it does not automatically license every voice or the words you supply. For this lab inspect nl_BE-rdh-medium at the pinned MODEL_CARD URL. Its dataset is listed as CC0 and its training section says trained from scratch, so there is no earlier fine-tuning base to chase in that inspected card. Preserve the card and its revision alongside the weights. A repository-wide MIT badge or an ONNX configuration is insufficient evidence of voice rights. Record what the documents actually say, not “free for everything.” Redistribution of engine code and voice weights are different decisions. Before a commercial integration or redistribution, check the applicable terms for that concrete use; this exercise supplies evidence rather than legal clearance.

If a required rights field is undocumented, write unresolved and name the missing source. Do not fill it from another voice’s card merely because the language or speaker label looks similar to your chosen artifact.

# Rights ledger fields; documentation only
engine = OHF-Voice/piper1-gpl; GPL-3.0-or-later
voice = nl_BE-rdh-medium
voice_revision = e7229f7d33a10e145e109b886d661984201ac3c4
dataset_in_card = CC0 1.0 Universal
training_in_card = trained from scratch
voice_rights_note = inspect exact card; no blanket grant from Piper

Budget hardware without promising speed

Begin with CPU execution on a supported Linux training machine; GPU acceleration is optional and introduces extra runtime and driver dependencies. Inventory architecture, available RAM, free disk and the package wheel supported by that system. Voice-file size is storage, not the full runtime memory requirement. Leave room for the environment, model pair, temporary files and final WAVs. The worked budget is deliberately bounded to one voice, two short texts and one process. If the machine cannot install a compatible wheel, stop at the documented setup report instead of renting a GPU or changing a production server. Do not promise real-time speech from a model label. Later you will measure initialization and end-to-end synthesis separately. The expected outcome here is an environment table and a clear “can test locally” or “blocked by compatibility/resources” decision.

Record whether the machine is already owned or rented and whether downloads are metered. No paid inference endpoint is required here, but that says nothing about electricity, storage, bandwidth or an unrelated hosting invoice.

# Read-only Linux inventory, no model execution
uname -m
free -h
df -h .

Practical assignment

Lab evidence

Continue to learning

Course outline

  1. Choose a private task and separate licenses
  2. Install an isolated engine and pinned voice
  3. Prepare text and specify a controlled first render
  4. Inspect audio structure, meaning and resource use
  5. Build a bounded two-file batch with explicit failure
  6. Package a reproducible private TTS project