Skip to content

Add Qwen Audio 3 milestone two controls - #5

Merged
Saco93 merged 5 commits into
mainfrom
feat/qwen-audio3-milestone2
Aug 5, 2026
Merged

Add Qwen Audio 3 milestone two controls#5
Saco93 merged 5 commits into
mainfrom
feat/qwen-audio3-milestone2

Conversation

@Saco93

@Saco93 Saco93 commented Aug 5, 2026

Copy link
Copy Markdown
Owner

Summary

Milestone 2 implementation and bounded authorized evaluation for Qwen-Audio-3:

  • add backward-compatible Standard, Low-latency dictation, Long-form, and Custom recognition presets
  • add documented multi_threshold_mode_enabled and optional speech_noise_threshold controls
  • parse bounded sentence/timed-unit metadata without changing transcript assembly
  • advance safe aggregate diagnostics to schema 4
  • add Regional/Custom endpoint routing for Beijing and Singapore
  • add optional hostname-based Workspace ID routing with transport-safe validation
  • preserve legacy/custom endpoints and keep endpoint, Workspace ID, models, credentials, transcript, and timed-unit text out of diagnostics

Evaluation

  • 17 fixed private WAV files replayed under three presets: 51/51 Streaming requests succeeded
  • all tested 250–2200 ms inserted-digital-silence samples retained both clauses
  • 15/15 speech files produced output under every preset; silence/noise-only false positives were 0/2 under every preset
  • one realtime session per preset completed; 73 timestamp-bearing results and 106 timed units were accepted with zero rejected/truncated units
  • one live cancellation canary produced no output
  • Beijing Regional empty-workspace Streaming and Native canaries succeeded
  • no corpus audio, references, transcripts, provider messages, credentials, endpoints, Workspace IDs, or raw diagnostics are committed

See docs/qwen-audio3-milestone2-evaluation.md for aggregate results and limits.

Scope boundaries

  • no timestamp identity-based deduplication or correction replacement
  • no reconnect/replay or retained Audio3 audio
  • no context collection or Filetrans
  • no automatic credential-region probing
  • no claim of Singapore feature parity or workspace-route success; those routes were not attempted without matching scoped credentials
  • Standard remains the default; bounded one-speaker results are not general accuracy or latency recommendations

Validation

  • cargo fmt --all -- --check
  • cargo check --locked --all-targets
  • cargo test --locked --all-targets — 241 passed
  • cargo clippy --locked --all-targets -- -D warnings
  • make validate-qml
  • git diff --check

Saco93 added 5 commits August 5, 2026 13:55
Record the official VAD fields, bounded timestamp semantics, regional endpoint
matrix, workspace routing, and credential scope before runtime implementation.
Block inferred threshold fields and identity-dependent timestamp deduplication
where the provider contract is insufficient.
Add backward-compatible standard, low-latency, long-form, and custom control
resolution for sentence boundaries. Expose the documented multi-threshold mode
and optional speech/noise threshold while preserving the standard request and
keeping candidate presets pending controlled live evaluation.
Parse sentence and timed-unit ranges with a bounded borrowed visitor that skips
private unit text, preserves transcript assembly on metadata rejection, and
emits numeric-only aggregate diagnostics. Advance diagnostics to schema 4 while
leaving identity-dependent deduplication and reconnect explicitly blocked.
Add presence-aware Regional and Custom routing for Beijing and Singapore, with
fixed Streaming and Native targets and optional hostname-safe workspace routing.
Preserve custom endpoints and dormant values while keeping route identifiers
out of errors and schema 4 diagnostics.
Record aggregate pause, noise, realtime timestamp, cancellation, and Beijing
routing results without including private corpus inputs or provider output.
Keep Standard as the default and preserve explicit limits on regional and
identity-dependent claims.
@Saco93
Saco93 merged commit e8a5120 into main Aug 5, 2026
2 checks passed
@Saco93
Saco93 deleted the feat/qwen-audio3-milestone2 branch August 5, 2026 10:13
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant