tacotron
Guides the legacy TensorFlow 1.x Tacotron repository for text normalization, speech-data preprocessing, Tacotron graph inspection, training, checkpoint synthesis, evaluation, and the Falcon demo server.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- tacotron
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Guides the legacy TensorFlow 1.x Tacotron repository for text normalization, speech-data preprocessing, Tacotron graph inspection, training, checkpoint synthesis, evaluation, and the Falcon demo server.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/installation.md2.2 KB
- 参考文档references/repo-provenance.md1.7 KB
- 参考文档references/repo-routing-metadata.json318 B
- 参考文档references/troubleshooting.md2.4 KB
03 SKILL.md 正文
Tacotron
Use this skill for the original keithito/tacotron-style implementation of
Tacotron speech synthesis in TensorFlow 1.x. It is a repository workflow skill,
not a modern TensorFlow 2 or Tacotron 2 guide. Keep the original model's
checkpoint, hyperparameters, text cleaners, and spectrogram conventions
consistent across preprocessing, training, and synthesis.
First route
- Text cleaning, number expansion, symbols, ARPAbet, or CMUDict: read
text-normalization. - LJ Speech/Blizzard layout, preprocessing, or custom data conversion: read
data-preparation. - Graph shapes, CBHG/prenet/attention internals, or audio dimensions: read
model-architecture. - Training commands, hparams, checkpoints, TensorBoard, or loss recovery:
read
training. - Checkpoint WAV synthesis, batch evaluation, or HTTP serving: read
synthesis-serving.
For a request spanning routes, use data-preparation → training → synthesis-serving, and use text-normalization whenever the cleaner or symbol contract changes.
Runtime contract
The repository is not packaged with setup.py, setup.cfg, or pyproject.toml.
Use an isolated legacy Python environment and run repository entry points from a
Tacotron checkout. The documented requirements omit TensorFlow; install a
TensorFlow 1.x build compatible with the platform. TensorFlow 1.15.5 with
Python 3.6 was verified for inspection, together with the pinned audio/text
dependencies, but newer Python/TensorFlow combinations are not automatically
compatible with tf.contrib.
Use explicit roots in command examples; replace these two assignments only when the checkout is stored elsewhere:
SKILL_ROOT=/path/to/tacotron-skill
CHECKOUT_ROOT=/path/to/tacotron-checkout
Bundled command builders run from SKILL_ROOT and print a command prefixed by
cd "$CHECKOUT_ROOT". Native preprocessing, training, evaluation, and serving
must therefore run with the checkout as their cwd; a skill directory is not a
source checkout. The skill's audio path uses the repository's librosa/SciPy
helpers and built-in Griffin-Lim reconstruction. It does not bundle or require
a neural vocoder, but real WAV input is required for preprocessing and a real
checkpoint is required for synthesis.
cd "$CHECKOUT_ROOT"
python -m pip install -r requirements.txt
# Select a platform-compatible TensorFlow 1.x package; do not use TensorFlow 2.x.
python -c "import tensorflow as tf; print(tf.__version__); print(hasattr(tf, 'contrib'))"
The repository's requirements pin old versions. If modern package resolution
breaks scipy, librosa, or numba, follow the compatibility notes in
references/installation.md rather than
blindly upgrading all dependencies.
Verification boundary
The bundled text, metadata, fixture, environment, shape, and command helpers are offline or dry-run checks. They do not prove real audio decoding, spectrogram quality, checkpoint compatibility, generated WAV quality, GPU performance, full preprocessing, training convergence, or server security. Full training and checkpoint synthesis need external data/checkpoints and can be expensive. Use the bundled builders and validators before starting them.
Shared invariants
- The default sample rate is 20 kHz; audio uses 80 mel channels, 1025 linear frequency bins, 50 ms windows, and 12.5 ms frame shifts.
outputs_per_step=5andmax_iters=200affect both training and synthesis; use the same relevant hparams at both times.- Preprocessing writes time-major
.npyspectrogram arrays andtrain.txtrows with two filenames, frame count, and text. Validate this before training. - Text conversion appends the
~EOS symbol. Cleaner names and ARPAbet syntax must be available in both training and inference environments. - Full training and checkpoint synthesis need real datasets/checkpoints and can be expensive. Use the bundled builders and validators before starting them.
Read references/troubleshooting.md for
cross-cutting TensorFlow-version, dependency, path, checkpoint, and audio
failure handling. Read references/repo-provenance.md
before deciding whether this skill is stale for a changed checkout.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.