Two new batch ASR models from moondream, both drop-in derivatives of Parakeet TDT v3 (same 25 languages, tokenizer and API). Existing apps are unaffected: .v3 stays the default.
Parakeet Ultra — AsrModelVersion.ultra (FluidInference/parakeet-ultra-coreml). A post-trained v3 that is more accurate everywhere at the same speed. int8 encoder (595 MB), Neural Engine by default, iOS 17+ / macOS 14+. Recommended for new integrations.
Parakeet Redux — AsrModelVersion.redux (FluidInference/parakeet-redux-coreml). A ternary v3 with a 183 MB 2-bit encoder (~220 MB download vs ~480 MB for v3). iOS 18+ / macOS 15+ only; on older systems it throws and points to .ultra. Neural Engine by default; the first ANE load compiles for several minutes (cached after), encoderComputeUnits: .cpuAndGPU skips it.
| Full sets, M-series Mac, ANE | v3 | Redux | Ultra |
|---|---|---|---|
| LibriSpeech test-clean WER | 2.27 % | 2.71 % | 2.13 % |
| LibriSpeech test-other WER | 4.12 % | 5.12 % | 3.81 % |
| FLEURS, 24 languages, mean WER | 14.81 % | 13.06 % | 11.67 % |
| test-clean RTFx | 128.6× | 83.9× | 126.7× |
let models = try await AsrModels.downloadAndLoad(version: .ultra)CLI: --model-version ultra|redux on transcribe, asr-benchmark, fleurs-benchmark (which also gains --encoder-compute-units), emission-delay and ctc-earnings. Docs: ParakeetUltra.md, ParakeetRedux.md.
What's Changed
- feat(asr): AsrModelVersion.redux — Parakeet Redux (2-bit ternary v3, iOS 18+) by @Alex-Wengg in #955
- feat(asr): AsrModelVersion.ultra — Parakeet Ultra (post-trained v3, int8 encoder) by @Alex-Wengg in #956
- docs(asr): refresh Redux/Ultra numbers on the v3-family long-form path by @Alex-Wengg in #957
Full Changelog: v0.17.2...v0.17.3