A language can change.
The engine should follow.

Nine more public FLEURS clips, three per language. Same M4 CPU and two-thread runtime; one model reused across the batch. Warm decode excludes load. Arabic WER and Japanese/Korean CER are shown; strict Unicode normalization retains diacritics and number differences.

Omni improves Arabic and Korean character scores here, while Dolphin is better on Japanese. Arabic word errors remain substantial; Korean spacing inflates word-level errors. Peak process memory reached 571 MB for Dolphin and 1.69 GB for Omni; loads were 0.212 and 0.481 seconds. No physical iPhone, conversational, or code-switch validation. These experiments do not enable new routes or establish a universal best model.

Nine public FLEURS clips, same three Mandarin, Japanese, and Korean clips. M4 CPU two threads, sherpa-onnx 1.13.8, INT8 ONNX export, automatic language and inverse text normalization enabled. This measures the CPU export, not FluidAudio Core ML or iPhone.

Mandarin zero errors covers only 81 reference characters, not general perfect accuracy. Japanese CER 8.02%; Korean CER 16.54% and whitespace WER 75%. Model load 0.538s; peak process memory 1.17 GB. Commercial-use clarification for the custom FunASR weight license is pending; no SenseVoice route is shipped or enabled. A maintainer discussion updated on 22 September is not a final licensing clearance.

One 60.24-minute job completed in the Mac Debug harness at e5387f11: the same 33.78-second English clip repeated 107 times. Ultra + Tiny; all transcription routes stayed on Parakeet, with no fallback ASR.

Debug build, single job, repeated source audio, uncontrolled host load. This checks long-job completion and resume, not independent accuracy, bulk queues, Release speed, or physical iPhone behavior. No WER/CER is assigned. Language detection remains the largest measured stage; avoiding fallback does not eliminate classifier cost.

Same Mac native services and public fixtures. Standalone runs include service scheduling/loading and then a repeated warm pass, without language detection. Auto v3 runs include language classification and selected Whisper fallback. Cache and execution order were not controlled.

On this 33.78-second English fixture, standalone v3 had 10.96% WER and Ultra 21.92%; segmented Auto Ultra had 4.11%. The English/German fixture favored standalone Ultra instead. This is not a global v3-versus-Ultra accuracy ranking. Standalone avoids routing overhead, but a faster transcript is not necessarily a more complete transcript.

Early native Mac pilot with Ultra primary and Small fallback. Tiny baseline plus Base/Small classifier runs; same public English, Hindi, mixed English–Hindi, and seven-language fixtures. These are full Auto timings, not detector-only speed.

Tiny remains the default. Hindi source-utterance language coverage was about 90.9% with Tiny, 35.9% with Base, and 32.4% with Small in these runs. Coverage includes internal utterance silence and counts unknown/mixed as not correct; it is not frame-annotated detector accuracy. Base’s first English run (120.72s) included cold compilation/loading: later cases must not be compared as an isolated model-speed ranking. Both larger classifiers failed on the seven-language fixture. Small sample and decoding errors prevent a general ranking.

All 14 cases completed on the iOS 27 arm64 simulator, source e5289bcf. Parakeet v3 + Tiny detection + Custom → Whisper Small. Seven monolingual recordings and seven stitched switches use the same hashed public corpus as the Mac batch. This is a Mac-hosted simulator with 24 GiB host memory, not a physical iPhone test. Hindi therefore passed the memory eligibility check.

App-reported elapsed times include routing and transcription. Separate model preparation took 74.27s; caches and host contention were not controlled. Mac results used Ultra, so these are not a direct platform speed comparison. Mixed-language CER includes all reference characters. Completion does not establish accuracy, thermal behavior, background execution, or memory safety on an iPhone.

41 of 42 cases completed; the seven-language Custom-policy case failed. M4 Mac, isolated native app compiled from e5387f11. Public FLEURS fixtures: English, Hindi, Mandarin, Arabic, Japanese, Korean, and German recordings plus seven stitched language-switch cases, each run with three fallback policies. Parakeet Ultra + Tiny classifier; selected Whisper Small fallback. Times include detection, local model load, transcription, and checkpoints, excluding app launch. New process and job per case; filesystem/Core ML caches not flushed.

In the English-only fixture, all three policies used Parakeet without a fallback transcription pass (3.34–3.83 seconds for 33.78 seconds of audio). This is a small Mac pilot, not an iPhone promise. Apple policy uses Apple only for eligible installed locales; other segments use Whisper. Custom policy uses only its enabled language/device routes and otherwise Whisper. Mixed-language CER includes English characters: it is not a Mandarin-only score. Sequential, contended execution and cache differences prevent an isolated speed ranking. Failure means no complete transcript and no invented error rate. No physical-iPhone claim.

• Initial batch: all 14 Custom cases failed at model setup because of a path bug; these had no ASR accuracy score.
• After the path fix (3290a8de): 12 of 14 Custom cases completed; standalone Mandarin and the seven-language case still failed.
• The next fix gave short speech fragments wider context without changing ownership boundaries. This final batch preserves all cases, including failures, instead of selecting only successful outputs.