Add Nemotron Unified ASR 0.6B (25 European languages) as an opt-in streaming model for German + English dictation #1682

What are you trying to do?

I dictate all day in German and English, often both in one dictation, on macOS (M4 Pro). Today I choose between live text and accuracy. Nemotron 3.5 streams, but it lost or garbled the opening of 6 of 33 of my dictations on Auto (#1517). Parakeet Ultra on Auto got all 33 right and is what I use now, but it doesn't stream.

NVIDIA released nvidia/nemotron-labs-unified-asr-0.6b on 27.10. It's the multilingual version of the Parakeet Unified English that Hex already ships: the same 0.6B unified FastConformer-RNNT, the same buffered streaming with 5.6 s left context, chunk and right context set from 2.08 s down to 0.16 s, punctuation and capitalization, and 25 European languages behind a language-ID prompt (target_lang="de", "en", … or "auto"). License OpenMDW-1.1, like Nemotron 3.5.

From its card, WER % with the language given:

                                 offline  1.12 s  0.56 s  0.32 s  0.16 s
German (FLEURS+MCV+MLS average)   4.32     5.45    6.30    7.47   10.01
English, 7 cleaned sets           4.68     5.41    5.68    6.27    7.87
Nemotron 3.5 streaming, English    -       7.26    7.58    7.97    8.90
Parakeet v3, English (offline)    4.93

FLEURS German alone is 3.87 offline and 5.24 at 1.12 s. Nemotron 3.5 streaming's card gives about 8.2 at 1.12 s. So at 1.12 s it streams German at about v3's offline level, and well ahead of Nemotron 3.5.

Mac ports already exist and look fast enough:

• FluidAudio CoreML: ~11–12 ms per 0.56 s step on the Neural Engine (M5 Pro).
• mlx-community: 27–69 ms per 80 ms step on an M1 Pro, 896 MB peak.

It isn't in transcribe.cpp yet. That's requested in handy-computer/transcribe.cpp#250.

What would help?

Nemotron Unified as an opt-in model in the picker, next to Parakeet Unified English, once transcribe.cpp carries it. Streaming for dictation, with a final pass on the same model.

One thing needs measuring before it's recommended for mixed-language dictation. The card's numbers are all with an explicit language. It warns that on auto at latencies of 0.56 s or less, "the first chunks do not contain enough speech" and it "can begin emitting tokens belonging to a different language", which can't be revised. That's the #1517 failure. It says the effect is "much smaller" at higher latency, but publishes no auto numbers. So:

• Default it to auto at 1.12 s (or 2.08 s) for users on Auto, not a low-latency setting.
• Check auto on German + English clips that switch mid-sentence, and on openings, against Parakeet Ultra's final pass. The fall back to the whole clip when a stream loses its opening or switches language #1520 fallback (re-transcribe when the live text loses its start or switches language) would catch the rest.

I can run it on my saved dictations (Keep recordings) as soon as there's a dev build, and report openings lost and German error against Ultra, as I did on #1517.