asrt-speech-recognition

Use this repo skill for tasks involving ASRT SpeechRecognition, a Chinese speech recognition system with ASRT-style WAV/data manifests, TensorFlow/Keras CNN+CTC acoustic models, a statistical pinyin-to-Chinese language model, and HTTP/gRPC serving clients.

Route by task

User task signal	Read
asrt_config.json, dict.txt, datalist/label format, /data/speech_data, WAV sample rate, spectrogram/MFCC/SpecAugment, DataLoader, CTC utility smoke checks	sub-skills/data-and-features/SKILL.md
SpeechModel251BN, Keras CTC model shapes, ModelSpeech, training/resume/evaluation, save_models, single-file acoustic prediction, TensorFlow/GPU expectations, PyTorch backend caveats	sub-skills/acoustic-models/SKILL.md
Decode ASRT pinyin such as ni3 hao3 ya5, ModelLanguage, language_model1.txt, language_model2.txt, streaming pinyin-to-text state, unknown pinyin diagnosis	sub-skills/language-model/SKILL.md
ASRT HTTP service, /speech, /language, /all, gRPC AsrtGrpcService, asrt.proto, payload construction, status codes, Docker/service deployment	sub-skills/serving-clients/SKILL.md

First operating checks

1. Read references/installation-and-runtime.md before installing dependencies or choosing CPU versus CUDA/GPU guidance.
2. Run the environment check when a user asks whether their runtime can operate ASRT modules:

   python scripts/check_asrt_runtime.py --help
   

3. Run the bundled self-contained smoke checks when validating this skill's helpers rather than a user's ASRT checkout:

   python scripts/run_asrt_smokes.py
   

4. Use references/troubleshooting.md for cross-cutting install/import, data/model-file, TensorFlow, and dependency failures before drilling into a sub-skill.

Cross-workflow recipes

Validate ASRT data before training

Route to data-and-features, validate config/list/dict consistency, then route to acoustic-models only after the data map, sample rate, and pinyin labels are clean. Do not start a training recipe when the user is still missing list files, label rows, pinyin dictionary entries, or 16 kHz WAVs.

Prepare an acoustic model run

Route to acoustic-models for model selection and weight naming. Use data-and-features for audio/features and language-model only after acoustic prediction has produced pinyin tokens. A CPU Keras construction smoke proves only model import/shape compatibility; it does not verify GPU training or accuracy.

Decode pinyin text only

If the user already has pinyin tokens such as ni3 hao3 ya5, skip acoustic/audio guidance and use language-model. Its bundled decoder is self-contained and includes ASRT language-model count files.

Build service clients

Route to serving-clients when the server is already prepared or the user is building a client. If startup fails because save_models/SpeechModel251bn.model.h5 is missing, route back to acoustic-models. If payload WAV metadata is wrong, route to data-and-features.

Boundaries and verification status

• This skill provides self-contained operating guidance and bundled helpers. It does not include trained acoustic weights or full speech corpora.
• Full training, evaluation accuracy, trained-weight prediction, live HTTP/gRPC round trips, and Docker builds require user-provided datasets/weights/services and are not claimed as verified by this skill alone.
• The generated language-model sub-skill bundles dict.txt plus language_model1.txt and language_model2.txt so pinyin-to-text decoding can be used without reopening the ASRT source tree.
• Repo provenance and routing metadata are in references/repo-provenance.md and references/repo-routing-metadata.json.