---

name: synthesis-serving description: "Guides Tacotron checkpoint loading, WAV synthesis, batch evaluation, and the Falcon demo server with matching hyperparameters and safe operational checks." disable-model-invocation: true metadata: disco-role: operating license: MIT

Synthesis and serving

Use this route for using a trained checkpoint, generating evaluation WAVs, starting the browser demo, or diagnosing empty audio and checkpoint errors. Read references/workflows.md before starting a server or changing hparams.

Workflow

1. Identify the exact checkpoint prefix, including its step suffix when using model.ckpt-<step>. Confirm sidecar checkpoint files and the hparams used to train it.
2. Use eval.py --checkpoint <prefix> for the built-in sentence batch, or demo_server.py --checkpoint <prefix> --port 9000 for the browser UI.
3. Pass the same relevant --hparams string used during training. Cleaner names, audio dimensions, outputs_per_step, and max_iters are not cosmetic.
4. Test the command with the bundled builder first. Keep the server bound and exposed deliberately; the original demo listens on 0.0.0.0.
5. Verify a generated WAV's sample rate and duration, and inspect endpoint trimming if output is silent or abruptly cut.

Command roots and synthesis boundary

Build from the skill root and run native evaluation/serving only from the checkout root. The checkpoint must be a real compatible prefix; it is not part of this skill.

SKILL_ROOT=/path/to/tacotron-skill
CHECKOUT_ROOT=/path/to/tacotron-checkout
cd "$SKILL_ROOT" && python sub-skills/synthesis-serving/scripts/build_synthesis_command.py --checkout-root "$CHECKOUT_ROOT" --mode eval --checkpoint /data/logs-tacotron/model.ckpt-185000

This repository reconstructs WAVs with its librosa/SciPy Griffin-Lim path; it does not use or bundle a neural vocoder. Real synthesis additionally requires the legacy TensorFlow/audio/Falcon dependencies as applicable, checkpoint sidecar files, writable output space, and matching hparams. The builder checks only command syntax: it loads no checkpoint, opens no port, and proves no audio quality, endpoint behavior, or server security. Full eval and server execution remain unverified without supplied weights and an approved local listener.

Read references/api-reference.md for Synthesizer lifecycle and endpoints, references/cli-reference.md for exact flags, and references/troubleshooting.md for checkpoint, hparam, text, and network-boundary failures. Use scripts/build_synthesis_command.py to build eval/server commands without loading a checkpoint or opening a port.