The pipeline targets Sichuan–Chongqing navigation speech. Normal samples use one decode; anomalous output triggers constrained retry, then legacy-model fallback if needed, followed by deterministic normalization and quality gates.
Navigation dialect ASR must handle accents and place names while avoiding loops, sentence repetition, abnormal length, and invalid characters.
Long GPU jobs may lose progress when a process exits. A single decoding strategy is not robust enough for anomalies, and retry or post-processing rules must not become an untraceable black box.
Each stage has a defined input, failure path, and persistent evidence.
Split audio deterministically for resume.
One greedy decode for normal samples.
Detect repetition, length, and character anomalies.
Constrained retry, then legacy model as needed.
Build JSONL, audit records, and SHA-256.
No further training in the current phase, preventing test output from feeding back into model or rules.
The legacy model is loaded only after retries fail, avoiding permanent dual-model GPU use.
Safe rules cover known navigation terms; experimental place-name rules remain isolated.
Final output is checked for UTF-8, JSONL, IDs, fields, and obvious generation anomalies.