Session Date/Time: 20 Mar 2025 06:00

mlcodec

Summary

The IETF 122 mlcodec working group session in Bangkok focused on progressing three key drafts: the Opus extension mechanism, Deep Redundancy (DRED), and speech coding enhancements. The session also included a proposal for adoption of a new scalable quality extension draft. Key outcomes included plans to proceed with working group last calls for the opus extension draft and continued refinement of DRED specifications.

Key Discussion Points

Opus Extension Draft

Frame Separator Issue: Tim presented a fix for handling frame separators with explicit increment of zero in the repeat extensions mechanism
The issue arose from discrepancy between specification text and implementation regarding which extensions get repeated
Proposed Solution: Frame separators with zero increment should be treated like padding, not updating the repeat pointer
Editorial improvements needed for clarity around extension vs. extension payload terminology
Decision made not to pursue complex extension numbering proposals due to excessive complexity

Deep Redundancy (DRED) Draft

Training Documentation: Jean-Marc added training procedure documentation to the Opus repository with scripts for dataset download and processing
Weight Standardization Discussion:
- Current weights are INT-8 quantized from FP-32 training
- Decision that implementers must use exact standardized weights (or quantizations), cannot train their own models
- Test vectors will be mandatory normative requirement in addition to weight specifications
Vocoder Standardization Challenge: Four options discussed for defining vocoder features:
1. Reference vocoder with deviation allowance
2. Objective metrics on vocoder synthesis
3. Invert vocoder to match decoded features
4. Standardize features independently
Jean-Marc plans to attempt option 3 (vocoder inversion approach)

Speech Coding Enhancements

No presentation due to Yon's unavailability
No substantive changes reported since becoming working group document
Ongoing work may result in future updates

Scalable Quality Extension (Individual Draft)

New Proposal: Jean-Marc presented draft for extending Opus beyond current limitations
Capabilities: Support for >8 bits depth per band, >20kHz bandwidth, up to 500 kbps per channel
Implementation: Uses extension mechanism with entropy-coded payload, maintains RFC 6716 compatibility
Technical Approach:
- PVQ refinement using upscaled codebook values
- Scalar parameter refinement for fine energy and angles
- Linear 2kHz bands above 20kHz (14 bands total for full 48kHz)
- New cubic codebook for high bands at high bit rates
Performance: Achieves -115dB THD+N at 480 kbps per channel
Backward Compatibility: Base layer identical to original Opus, passes existing test vectors

Decisions and Action Items

Opus Extension Draft

Action: Tim to publish update implementing frame separator fix and editorial improvements
Timeline: One month (mid-April)
Decision: Proceed with working group last call after Tim's update

DRED Draft

Action: Jean-Marc to develop test vectors as normative requirement
Action: Add training procedure documentation to draft (not just repository)
Action: Specify both FP-32 original weights and INT-8 quantized weights
Action: Resolve vocoder standardization approach before next meeting
Decision: Both standardized weights AND test vector compliance will be normative requirements

Scalable Quality Extension

Decision: Chairs will initiate working group adoption call on mailing list
Action: Working group members to review draft and provide feedback on adoption

Next Steps

Immediate: Working group adoption call for scalable quality extension draft
Mid-April: Updated opus extension draft publication followed by working group last call
Next Meeting: DRED vocoder standardization resolution and potential updates to speech coding enhancements
General: All documents significantly behind milestone schedule - acceleration needed before Madrid meeting

The working group emphasized the need to accelerate progress on all work items to meet charter obligations and prepare documents for IESG review.