mirror of
https://github.com/FluidInference/FluidAudio.git
synced 2026-06-11 20:24:36 +00:00
AlexandGitHub
cbbabf6870
The experimental Parakeet CTC 0.6B zh-CN model (8.23% mean CER, 14.8x RTFx on THCHS-30) is dominated by two in-house Mandarin models: Paraformer-large (2.12% CER, 85x RTFx) and SenseVoice (3.09% CER, 382x RTFx). There is no remaining use case where the CTC zh-CN model is the best choice. Removed: - CtcZhCnManager / CtcZhCnModels (Sources/FluidAudio/ASR/Parakeet/SlidingWindow/CTC/) - ctc-zh-cn-transcribe and ctc-zh-cn-benchmark CLI commands + help text - .ctcZhCn AsrModelVersion case and the isCtcOnly guard machinery - Repo.parakeetCtcZhCn / ModelNames.CTCZhCn entries - THCHS-30 dataset plumbing (ChineseDatasetDownloader, exclusively used by the zh-CN benchmark) - CtcZhCnTests and the CTC-only validation tests in AsrModelsTests - zh-CN benchmark leg in Scripts/parakeet_subset_benchmark.sh - Documentation: Benchmarks.md zh-CN section, Models.md rows, DirectoryStructure.md entries, README highlight Kept (shared with the English CTC 0.6B/110M keyword-spotting path): CtcDecoder.swift, ARPALanguageModel.swift, and all CTC 110M / ctc-earnings-benchmark code. The Japanese hybrid TDT model (parakeet-ctc-0.6b-ja repo naming notwithstanding) is untouched. Capability gap (follow-up, not blocking): the CTC zh-CN manager had built-in 15s sliding-window long-form support; the Paraformer and SenseVoice hosts do not yet chunk long audio. Tracked as a follow-up. ### Why is this change needed? <!-- Explain the motivation for this change. What problem does it solve? -->