German speech recognition: tiny Parakeet keeps pace with Whisper, Turbo falls behind

What this is about

An unusual number of speech recognition models landed at once this week. Salesforce released SOV-Transcribe-1, Elevenlabs shipped Scribe v3, and two weeks earlier Nvidia put out Nemotron 3.5 Transcribe. Then there are offline dictation apps like ThumbFree, built on Nvidia's small Parakeet model. Nearly all of them advertise their scores on Artificial Analysis's AA-WER index. By its own description, that index only measures English datasets: conversations with voice agents, speeches from the European Parliament and earnings calls.

So we collected the German numbers: the German results of Hugging Face's Open ASR Leaderboard, a dialect study from Austria and a German medical study. Everything is measured in word error rate (WER), the share of words that are misrecognized, dropped or made up. 4% means a 1,000-word dictation contains about 40 errors.

The German ranking: small beats big

Here's the current German ranking from the leaderboard, as mirrored by Papers with Code (excerpt):

Technically, first place goes to "Sophea ASR K1" at 2.54%. But the accompanying paper describes a Greek-English recognizer with no released weights and never mentions German. We think the entry is a mislabeling and have left it out.

Parakeet v3 stands out most. At 0.6 billion parameters, it's less than half the size of Whisper large-v3 (1.55 billion) and still ties it. According to the leaderboard paper, its RTFx of 1720 (the seconds of audio it processes per second of compute) makes it roughly 15 times faster than Whisper at 111. The popular Whisper Turbo, on the other hand, lands at 6.12%, about 50% more errors than full-size Whisper. Plenty of desktop apps ship exactly this variant because it's faster and uses less memory. OpenAI also says Turbo can't translate.

The numbers depend heavily on when the evaluation was run. In the leaderboard paper from March 2026, Whisper was still at 4.26%, Parakeet at 4.20% and Voxtral Small at 3.01%. Two small independent tests back up the trend. A test on dev.to using 100 German FLEURS sentences on an M2 Pro Mac found 4.8% for Parakeet, 5.1% for Whisper Turbo and 6.9% for Apple's built-in speech recognition. Note, though, that the author builds an app on Parakeet himself.

What the new models show on German (and what they don't)

SOV-Transcribe-1 supports seven languages including German, according to its Model Card, and is licensed under Apache 2.0. But the benchmark table is only available as an image, and the model doesn't appear in the German leaderboard. BenchLM only lists an average of 4.88% across six European languages, so there's no German-only score. According to OrcaRouter, the BF16 checkpoint is around 53 GB, too big for a 48 GB card. No quantized versions exist yet. Quantization means storing the model's numbers at lower precision so it needs less memory. Salesforce tested on an RTX PRO 6000 with 96 GB. The model can't stream, meaning it can't produce text while someone is still talking.

Nemotron 3.5 Transcribe is a mixture-of-experts model: of its 30 billion parameters, only about 4 billion are active per step. For German, Nvidia reports 2.73% on FLEURS and 2.99% on Common Voice. The Model Card targets an H100 with 80 GB. An FP8 version runs on the 48 GB L40S, and the NVFP4 version on the RTX 5090. This model also only transcribes recordings, not live audio. Nvidia says the OpenMDW-1.1 license allows commercial use.

Scribe v3 isn't open and only runs through Elevenlabs' API. A report from all-ai.de, citing Elevenlabs, says German WER dropped from 3.7% (v2) to 2.1%. In its own blog, Elevenlabs only claims "42% better," measured on an internal version of FLEURS. That's a good illustration of how far measurements diverge: the same Scribe v2 scored 2.27% in the leaderboard paper, not 3.7%.

Hy ASR 3.0 doesn't make our cut. According to its technical report, it handles Mandarin, English and Chinese dialects, and no weights are linked there.

Why benchmark numbers can mislead

The leaderboard measures on FLEURS and Common Voice, which means read-aloud, cleanly recorded sentences. Before scoring, punctuation and capitalization are stripped using normalization modeled on English. So whether the formal "Sie" or the lowercase "sie" was written correctly doesn't count.

The Vienna-based company CASABLANCA looked into this systematically. Normalization alone shifted Whisper large-v3's WER by up to five percentage points. Then there's the host: on Artificial Analysis, the same Whisper large-v3 scores 4.1% on English at fal.ai, 4.5% at Together AI and 10.1% at Replicate. Hosted services rarely disclose their settings, so the authors say the gap can't be fully explained. On dialect data from Germany, Austria and Switzerland, error rates rose sharply. Voxtral Mini went from 3.8% on standard German to 6.4% on dialect, Deepgram Nova-3 from 5.2% to 8.3%. The study's table is sloppily formatted in places. For Parakeet, for example, it apparently compares v3 (Together AI) with v2 (Nvidia), which Artificial Analysis lists as separate models.

It gets even clearer in a study by Pforzheim University and XPACE using real German patient intake interviews, complete with medical terminology, an accent and a strong dialect:

Whisper makes three times as many errors here as on FLEURS. The authors themselves point out inconsistencies in model naming (Whisper is labeled "Distilled" once in the table). Parakeet wasn't tested. Only the cloud services delivered speaker attribution (diarization, i.e. who said what) out of the box. Open models needed add-on tools like pyannote.

Streaming: the price of live text

If you need text in real time, say for captions or phone bots, you'll give up accuracy. Voxtral Mini 4B Realtime shows this directly. At 160 ms latency, it scores 9.50% on German FLEURS, 6.19% at the recommended 480 ms and 4.15% at 2.4 seconds. Nemotron 3.5 ASR Streaming 0.6B lands between 8.3 and 9.8% depending on chunk size. A hands-on review by wavect adds that it lacks diarization, timestamps and custom vocabulary. Qwen3-ASR can stream via vLLM, which raises its average WER from 2.69 to 3.33%, according to the Model Card.

The big overview

For Parakeet, the WER applies to regular transcription of recordings. Cohere's Transcribe model (2B parameters, Apache 2.0) scored 3.84% in the March paper but isn't in the current list. Downloading it requires approval.

Our verdict

Dictation on your phone or laptop: In our view, Parakeet v3 is currently the best choice. It recognizes German as well as full-size Whisper but runs on the CPU. You can use it via ThumbFree (pick the multilingual model in the menu) or the desktop app Handy. One catch: the model detects the language on its own and can't be locked to German. In the Hugging Face discussions, users complain about stray words from other languages and swallowed sentences. If an app offers Whisper Turbo, pick Parakeet or the full large-v3 for German instead.

Meeting minutes: Accuracy matters more than speed here. With a 24 GB card, Qwen3-ASR-1.7B or Voxtral Mini 3B are realistic options. For companies with server GPUs, Voxtral Small and Nemotron 3.5 Transcribe play in the top league, while SOV-Transcribe-1 still has to prove itself on German. With open models, you'll usually have to add speaker attribution yourself.

Call centers: Live text costs noticeable accuracy on German. Voxtral Realtime at 480 ms is the best open compromise. 8 kHz phone audio, dialects and jargon make everything worse, and no public leaderboard tests for that. The medical study and the dialect data show how quickly two to three percent turns into seven to twelve. Our recommendation: before you deploy, have an hour of your own audio transcribed by hand and measure for yourself. That tells you more than any leaderboard, ours included.