feat(cohere): Add Cohere Transcribe CoreML conversion with critical fixes#41

Draft

Alex-Wengg wants to merge 43 commits intomainfrom

docs/cohere-transcribe-coreml-decoder-fix

Commits on Apr 6, 2026

fix(cohere): Implement stateless decoder to fix cache repetition bug
Alex-Wengg
committed
docs(cohere): Update README with stateless decoder status and complete file organization
Alex-Wengg
committed
chore(cohere): Remove broken/experimental export scripts
Alex-Wengg
committed
chore(cohere): Delete archive and mark broken HF uploads
Alex-Wengg
committed
refactor(cohere): Clean up test suite and add PyTorch baseline
Alex-Wengg
committed
chore(cohere): Delete redundant utility scripts
Alex-Wengg
committed
feat(cohere): implement stateful decoder with Qwen3 approach

Alex-Wengg
and
claude
committed
test(cohere): add comprehensive benchmarks for stateful decoder

Alex-Wengg
and
claude
committed
chore: remove outdated debug scripts, logs, and reference code
Alex-Wengg
committed
feat(cohere): add 256-token decoder and investigation scripts

Alex-Wengg
and
claude
committed
docs(cohere): Identify encoder as root cause of quality issues

Alex-Wengg
and
claude
committed
docs(cohere): Complete root cause analysis - encoder training data bias

Alex-Wengg
and
claude
committed
fix(cohere): Correct audio window to 35 seconds (3500 frames)

Alex-Wengg
and
claude
committed
docs(cohere): Document .mlpackage requirement and .mlmodelc limitations
Alex-Wengg
committed
docs(cohere): Update README with current status and .mlpackage requirement
Alex-Wengg
committed
chore(cohere): Clean up obsolete files and failed experiments
Alex-Wengg
committed
chore(cohere): Remove remaining obsolete files
Alex-Wengg
committed
chore(cohere): Remove temporary upload docs and obsolete tests
Alex-Wengg
committed
chore(cohere): Remove obsolete build artifacts and test files
Alex-Wengg
committed
refactor(cohere): Organize original PyTorch files into cohere-pytorch subdirectory
Alex-Wengg
committed
docs(cohere): Add historical context and verified performance results
Alex-Wengg
committed
feat(cohere): Add INT8 quantized models and benchmarks

Alex-Wengg
and
claude
committed
refactor(cohere): Reorganize scripts and create unified benchmark tool

Alex-Wengg
and
claude
committed
refactor(cohere): Use jiwer library for text normalization

Alex-Wengg
and
claude
committed
fix(cohere): Use google/fleurs dataset with correct field names
Alex-Wengg
committed
refactor(cohere): Organize test files and scripts
Alex-Wengg
committed
docs(cohere): Add comprehensive research analysis and limitations
Alex-Wengg
committed
feat(cohere): Add stateless decoder variant (Parakeet approach)
Alex-Wengg
committed

Commits on Apr 7, 2026

Commits on Apr 9, 2026

research: Comprehensive investigation of Cohere multilingual ASR failure
Alex-Wengg
committed