Run audio transcribe stacks from any directory, without an input or installed models. It prints JSON with one stacks array. Each row contains id, characterization, native, requires_add_on, and impossible, projected from the same declared stack table used by request resolution. No runtime probe or readiness claim is part of discovery. audio transcribe capabilities --help leads to this list and describes how to query one choice for an original input. Missing-stack and unsupported-capability requests retain their existing semantic JSON refusals.
Finished transcription model stages also retain structured responses and JSON requests when used; alignment diagnostics explains how a saved alignment_unavailable entry joins the affected segment to its raw unit response. Raw streams and structured responses are separate evidence.
audio transcribe run prints host stage start/finish notices on stderr and a child elapsed-time notice every ten seconds while a child remains running. Elapsed time is liveness information; it is neither a percentage nor a prediction of remaining time. A child exit notice reports the process exit code, while the host still validates its result protocol before reporting stage completion. An exit-zero child with malformed result JSON therefore remains a failure.
Before each decode or model child launches, the host announces absolute paths to stdout.log and stderr.log. The default is a fresh private audio-transcribe-* temporary directory, retained until operating-system or user cleanup. The child writes its original bytes directly to those files, including carriage returns, escape sequences, and non-UTF-8 bytes. Logs are available during execution and retained after success, failure, or handled interruption. The transport also accepts a caller-selected log root for durable storage, with a unique run directory and separate numbered stage directories; see the integration API. An invalid or unwritable requested root fails before child execution, with no temporary fallback. There is no automatic retention policy or transcript-schema change for log locations.
The host does not classify or filter backend warning text. Backend stdout is preserved alongside stderr; neither stream is mixed into canonical transcript text or machine stdout. Treat warnings as evidence to investigate, not proof of harmlessness, recognition quality, or failure. Use exit status, semantic refusals, and result coverage for command outcomes. The decoded diagnostic tail may still appear in a failure's existing detail field; the retained files are the byte-exact evidence.
For VibeVoice startup messages about a missing preprocessor configuration, tokenizer class mismatch, or deprecated torch_dtype, the pinned-source diagnostic note identifies the relevant provider locations. The stage's processor/model calls pass the managed tokenizer directory, local_files_only=True, and dtype=torch.bfloat16; its comments tie these messages to provider commit 94da20d98b2fa7688e9cbfaf7692ddb4954f7600. This establishes configuration provenance, not a general warning verdict. Read the announced raw stderr.log alongside stdout.log and the run outcome; other warning text still needs its own evidence.
Keyboard interruption terminates and reaps the direct child before closing the log handles; an unresponsive child is killed after five seconds. Bytes already written survive; bytes a backend never flushed from its own userspace buffers cannot be recovered. Abrupt host termination such as SIGKILL cannot run Python cleanup, though bytes already written to the retained files remain. A host error creating or reading diagnostic files becomes a transport failure.
audio enhance renders stage lifecycle events on stderr for preparation, decode, inspection, stage evaluation, normalization, encoding, encoded-output checks, and publication, with elapsed heartbeats while synchronous work is blocked. Dry runs identify themselves and finish after prediction without encoding or publication. A stage finishing means its evaluation returned; skipped, no-op, applied, and abstained decisions remain in the report. Progress and logging describes the callback and context lifetime; the pipeline itself does not print or start a timer.
audio report summary /absolute/path/output.wav.report.jsonThis stdout-only JSON projection reads the saved audio_enhancement_report schema version "1". It does not open the source media, invoke detection, run models, remeasure audio, or rewrite either the report or transcript. Invalid JSON, duplicate keys, nonfinite numbers, wrong report kinds, unknown schema versions, and malformed required structures refuse at exit 2 using the enhancement commands' existing JSON error envelope.
| Summary field | Meaning |
|---|---|
kind, schema_version |
audio_enhancement_summary, "1"; identifies the projection |
report, source, profile |
Absolute report path, recorded source path/hash when present, and profile name/version |
rendered, dry_run |
Copied run state; a render is not proof that every requested goal was met |
stages |
Recorded stage name/status/reason, full component and final-region evaluations when present, and report pointers; bulky operations remain in the original report |
unresolved |
Full recorded unresolved rows with nested scope/evidence intact, plus pointers; absent in the summary when absent in the source report |
measurement_scopes |
Pointers to available before, predicted, and after blocks and their program_actual/regional blocks when present; no copies of those measurement blocks |
region_basis, timeline_preserved |
Copied when present, retaining the original report's scope and limits |
final_peak_validation |
Recorded status when present, plus a pointer to full validation details |
report_pointer is a JSON Pointer into the file named by report; for example, /unresolved/0 identifies one complete unresolved record and /measurements/after/program_actual identifies delivered-file program measurements. Before and predicted blocks never acquire an after label. Per-decision measured_at stays absent when the report did not supply it; a rendered report does not retroactively make every stage decision a post-encode measurement. Older reports without program_actual retain a pointer to the containing measurement block and acquire no replacement measurement.
Unknown additional evidence keys inside components and unresolved rows are retained, including structured noise_reference_scopes; required known structures are validated, and unrelated top-level report extensions remain accessible through the original file. applied remains the parent's operation status even when a child abstains. Skipped/no-op stages remain visible. There is no overall success enum: an empty unresolved ledger describes recorded checks and cannot establish perceptual quality, completeness of detection, or achievement of unmeasured goals.