---

name: serving-clients description: "Deploy and integrate ASRT HTTP/gRPC serving clients once trained weights are available." metadata: disco-role: operating disable-model-invocation: true license: GPL 3.0

Serving Clients

Use this sub-skill when the ASRT service boundary is the active problem: starting an already prepared HTTP/gRPC server, constructing request payloads, calling the service, interpreting responses, or debugging deployment and client integration failures.

Route here

• Start or configure ASRT HTTP/gRPC service processes after trained acoustic weights and language-model assets are already present.
• Build HTTP JSON payloads for GET /, POST /, POST /speech, POST /language, and POST /all.
• Integrate gRPC clients for Speech, Language, All, and Stream using the AsrtGrpcService contract.
• Interpret ASRT response/status fields, including streaming partial status 206000 and final status 200000.
• Debug endpoint paths, ports, base64 encoding, WAV metadata, protobuf stubs, and Docker CPU-only serving behavior.

Not here

• Training, exporting, or selecting acoustic model weights.
• Dataset preparation or audio feature extraction beyond reading WAV metadata and sample bytes for service payloads.
• Language-model corpus construction, scoring internals, or pinyin-to-text quality tuning.
• Copying or rewriting ASRT server implementations.

Read first

• Deployment boundaries and startup
• HTTP JSON API
• gRPC API
• Troubleshooting

Bundled runtime helpers

• scripts/make_http_payload.py builds JSON request bodies from a user-provided WAV file or pinyin sequence without importing ASRT.
• scripts/asrt_http_client.py calls the ASRT HTTP service with only Python standard-library networking and WAV handling.

Safe help checks:

python scripts/make_http_payload.py --help
python scripts/asrt_http_client.py --help

Example payload/client flow:

python scripts/make_http_payload.py --endpoint /all --wav sample.wav --pretty > all-payload.json
python scripts/asrt_http_client.py all --wav sample.wav --base-url http://127.0.0.1:20001
python scripts/asrt_http_client.py language --sequence-pinyin ni3 hao3 ya5

Critical operating facts

• The original HTTP and gRPC server scripts instantiate SpeechModel251BN, load save_models/SpeechModel251bn.model.h5, and load model_language at module import time before serving requests.
• Those server scripts start long-lived listeners (waitress.serve for HTTP; grpc.server(...).start() plus an infinite sleep loop for gRPC). This generated runtime therefore bundles only client and payload helpers, not server copies.
• HTTP audio requests send raw WAV sample frames, not the RIFF/WAVE container, as URL-safe base64 text in samples with matching sample_rate, channels, and byte_width metadata.
• HTTP language requests send sequence_pinyin as a JSON list of pinyin syllable strings.
• gRPC audio requests send the same raw sample frames in WavData.samples; gRPC language requests send repeated pinyins in LanguageRequest.

Evidence base

This sub-skill was revised from the serving evidence in asrserver_http.py, asrserver_grpc.py, client_http.py, client_grpc.py, assets/asrt.proto, Dockerfile, README.md, README_EN.md, and utils/ops.py.

Routing notes

• Missing save_models/SpeechModel251bn.model.h5 is a model-preparation problem; stop server/client debugging and route to acoustic-model work.
• Invalid WAV sample shape, sample rate, channel count, or sample width belongs to data/audio preparation after confirming the client payload mirrors the WAV metadata.
• Bad pinyin-to-text output with successful /language or gRPC Language responses belongs to language-model work.
• For protobuf, generated-stub, version, or service-method mismatches, use gRPC API and Troubleshooting before changing application logic.