German speech recognition locally: the most accurate models, dialects and company jargon

Which speech recognition model is most accurate for German

On published German test figures, Mistral's Voxtral Small 24B has the lowest word error rate of the open models in October 2026: 3.29 per cent in the Handy benchmark, ahead of OpenAI's Whisper large-v3 at 4.13 per cent. The other leading open models, among them NVIDIA's Parakeet and Canary, lie between about 4.2 and 5.3 per cent on the same read sentences. On your own recordings, dialect, names, company terms and room acoustics change the result more than the choice among these models, and the most accurate one needs far more GPU memory than Whisper.

This comparison is for anyone whose German transcripts are not yet good enough: a managing director whose meeting minutes come out wrong, a quality manager with technical terms, an IT team choosing a model for a transcription service. The word error rate, WER for short, counts the words a model gets wrong, adds or leaves out, per 100 words of a correct reference transcript. A WER of 4 per cent means about one word in 25 needs correcting. Model sizes and speed for servers are in our guide to Whisper GPU requirements; here the subject is accuracy in German and what you can do about the errors.

German results in the published rankings

The Handy benchmark is published by the makers of the dictation app Handy for their transcribe.cpp engine. It runs every model on the test split of FLEURS, a public set of read sentences, with weights quantised to Q8_0, meaning stored with 8 bits per weight. Its German table lists 27 models; these are the ones that matter for a company decision.

Handy German model ranking (models.handy.computer/languages/de), FLEURS test split, Q8_0, read on 30 October 2026; licences as listed there. OpenAI's own Whisper repository states the MIT licence for Whisper's code and weights.

The Open ASR Leaderboard paper (arXiv 2510.06961v3, 10 December 2025), from Hugging Face, the University of Cambridge and Mistral AI, measured differently. Its German figure averages the CoVoST-2 and FLEURS test sets, and it removes punctuation and casing and applies an English normalisation pipeline "closely following Whisper" before counting errors.

Open ASR Leaderboard paper, arXiv 2510.06961v3, Table 4 (multilingual), leaderboard data as of 8 October 2025, read on 30 October 2026. Voxtral Small 24B is not in that table.

NVIDIA's own model card for Parakeet TDT 0.6B v3 gives 5.04 per cent on German FLEURS and 4.84 on CoVoST. The same model therefore appears at 4.90, 5.04 and 5.24 per cent depending on who measured it, and Whisper large-v3 at 4.13 or 4.97. For the streaming model Nemotron 3.5 ASR, NVIDIA's card gives 8.22 to 9.90 per cent on German FLEURS depending on the chunk size, against 10.33 in the Handy table. Quantisation, normalisation, streaming settings and the choice of test sets each move the figures. Read the rankings in tiers: Voxtral Small ahead, then a cluster of Whisper large-v3, Qwen3-ASR, Canary and Parakeet within about one point, and the small and streaming models behind.

What the leading models need and where they stop

Hugging Face model cards of Voxtral Small 24B 2507, Whisper large-v3, Qwen3-ASR-1.7B, Canary-1B v2 and Parakeet TDT 0.6B v3, Cohere's release blog of 26 March 2026, OpenAI's Whisper repository for the 10 GB figure, file sizes at Q8_0 from the Handy table; all read on 30 October 2026.

Voxtral Small 24B supports English, Spanish, French, Portuguese, Hindi, German, Dutch and Italian under Apache 2.0. Its card states that it "requires ~55 GB of GPU RAM in bf16 or fp16" and "handles audios up to 30 minutes for transcription", so longer recordings are cut into pieces. It runs in vLLM 0.10.0 or later and detects the language by itself. Whisper large-v3 has 1,550 million parameters, and its card warns that the models "exhibit disparate performance on different accents and dialects of particular languages" and that predictions "may include texts that are not actually spoken in the audio input". Qwen3-ASR-1.7B covers 30 languages, German among them; its dialect list names only Chinese dialects. Cohere Transcribe, a 2B model, "is trained to expect a language tag and monolingual audio", which matters for meetings that switch between German and English. Parakeet detects the language automatically and is by far the fastest of these models: 3,333 seconds of audio per second on one A100 in the leaderboard paper, against 145.5 for Whisper large-v3.

Mistral's announcement of the Voxtral 3 family, with benchmark notes dated 20 October 2026, adds context biasing, where "key terms and free-form text can be given to bias the transcription and improve recall on technical terms", and speaker identification in Voxtral Small Transcribe 3. As of 30 October 2026 the announcement offers "self-deployment on your own infrastructure" but states neither open weights nor a licence, and Mistral's Hugging Face page lists no Voxtral 3 model, so we do not plan a local installation on it until both are published.

Tell us what you record and in which dialects, and our team proposes the models and the card for them.

Transcribing Swiss German, Bavarian and Austrian dialects

Dialect is where German speech recognition loses most accuracy, and the research gives a consistent picture: the large models understand much of the dialect and write it down in Standard German.

Swiss German. A University of Zurich study (arXiv 2404.19310, May 2024) tested Whisper large-v3 without any extra training. Against Standard German reference texts it reached 23 per cent WER on the STT4SG-350 and SwissDial test sets, 29.5 per cent on the Swiss Parliaments Corpus and 37.2 per cent on mock clinical interviews; a wav2vec2 model fine-tuned on Swiss German reached 14.0 per cent on STT4SG-350. The authors point out that Whisper usually outputs a Standard German translation, which the metric partly penalises, because Swiss German has no standard written form. Human raters scored meaning at 4.36 and fluency at 4.39 out of 5. Four of sixteen interview transcripts contained invented sentences, and the qualitative analysis found errors with names and numbers. Their conclusion: Whisper is usable for Swiss German "so long as the Standard German output is desired".

Bavarian, Franconian and Swabian. A study by LMU Munich and Bayerischer Rundfunk (arXiv 2506.02894, 2025) built a test set from about four hours of dialect speech, read children's stories from the broadcaster. Whisper large-v3 was the best model and reached 31 per cent WER against Standard German references on the dialect audio, compared with 9 per cent on the set's Standard German recordings. The authors found no consistent differences between regions and frequent errors on dialect-specific words. Fine-tuning on Standard German data gave mixed results.

Austria. The Bavarian recordings in that study come from regions in Bavaria. For Austrian German, a 2024 study by the Know-Center and Graz University of Technology (KONVENS 2024) used casual conversations without a set topic between speakers who know each other well, from the Graz Corpus of Read and Spontaneous Speech: Whisper large-v2 without extra training reached 41.8 per cent WER, and a wav2vec2 model fine-tuned on the same corpus, with a lexicon and a language model, 22.8 per cent. Spontaneous conversation is a much harder case than read sentences, so for Austrian speakers the test on your own recordings described below remains the reliable measure.

Fine-tuning on dialect audio. ZHAW's study of Swiss radio and TV speech (arXiv 2506.08836, June 2025) trained Whisper on about 230 hours of matching broadcast audio. Whisper large-v3 improved from 21.1 to 17.1 per cent WER and large-v3-turbo from 26.5 to 18.7. Training Whisper medium only on a set of read sentences made it worse on broadcast speech, and the remaining errors were mostly uncommon names of places, brands and people. A May 2026 preprint on the All Swiss German Dialects Test Set (arXiv 2606.07608) reports a smaller gain, from 28.56 to 25.60 per cent with a full fine-tune, and argues that much of the remaining error count reflects valid differences in wording. A fine-tune pays off when you have many hours of your own audio with corrected transcripts.

Company jargon, names and numbers

Product names, people, part numbers and abbreviations are the errors users notice first, and every study above found them among the remaining mistakes. Four tools help, in order of effort.

• A prompt and hotwords. faster-whisper, the runtime under many Whisper apps and compared in our guide to faster-whisper, WhisperX and whisper.cpp, accepts an initial_prompt, "optional text ... to provide as a prompt for the first window", and hotwords, "Hotwords/hint phrases to provide the model with". A list of your product names, customer names and abbreviations goes there.
• A fixed language. Its language parameter takes a code such as "de"; without it the language is detected from the start of the audio. For meetings that mix German and English, setting German avoids a wrong guess, and Cohere Transcribe needs the tag anyway.
• A glossary check after transcription. A local language model compares the transcript with your glossary and marks or replaces terms that were heard wrong, with a person approving the changes. This needs no change to the speech model. Keep the task that narrow: in the Graz study, a language model asked to correct whole transcripts raised the WER for most of the utterances tested.
• Fine-tuning on your own recordings with corrected transcripts, as in the ZHAW study, for dialect regions or a vocabulary that the first three steps do not fix.

Voxtral Small's card states that system prompts "are not yet supported", so the prompt route is not available for it. The context biasing of Voxtral 3 would cover this case once the weights and a licence are published.

Testing on your own recordings before you choose a model

A test with your own audio takes a few hours and tells you more than any ranking. Choose 20 to 30 minutes that cover your speakers, dialects, rooms and microphones. Write the reference transcript by hand, in the form you want: Standard German for minutes, or close to the spoken words for interviews. Bring the model output and the reference into the same form, with the same case, punctuation and way of writing numbers, then count substitutions, deletions and insertions and divide by the number of words in the reference. Count names, terms and numbers separately, since those decide whether a transcript is usable. Run two or three models, for example Whisper large-v3, Parakeet and Voxtral Small, and listen to every passage where a model wrote text that nobody said.

For this test you can order the workstation alone, the setup of Whisper, Parakeet and Voxtral Small alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Which GPU each German model needs

The ladder below starts with the smallest configuration that does the job. Usable memory follows our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB.

Memory from NVIDIA's product pages as listed on our GPU page; usable 18.5, 61.5 and 83.0 GiB by our rule; model memory from the cards cited above; our estimate.

Entry. The RTX PRO 4000 Blackwell leaves 18.5 GiB by our rule, enough for Whisper large-v3 at about 10 GB or any of the smaller models, and for Voxtral Mini 4B Realtime, which by its card "can run on a single GPU with >= 16GB memory". Live transcription with that model is covered in our guide to live captions on-premise.

Accuracy. The Handy figure comes from an 8-bit file of 24.04 GB in transcribe.cpp; the 55 GB, which are 51.2 GiB, apply to serving in BF16 with vLLM, the route the model card documents. The RTX PRO 5000 Blackwell with 72 GB leaves 61.5 GiB by our rule, about 10 GiB more than the model needs; the 48 GB version, with 40.0 GiB, is too small for it in BF16. We found no published throughput for Voxtral Small on RTX PRO cards, so how many hours of audio it handles per day depends on the recording length and on how vLLM batches the requests; a test run gives the figure for your audio.

Comparison and fine-tuning. One RTX PRO 6000 Blackwell Workstation Edition or Max-Q with 96 GB leaves 83.0 GiB, room for Voxtral Small and Whisper large-v3 side by side. ZHAW fine-tuned Whisper on two A100 cards with 40 GB each; one RTX PRO 6000 has more memory than both together, and the training time depends on the hours of audio and the number of passes, for which we found no published figure on RTX PRO cards. Two RTX PRO 6000 Workstation Edition cards at 600 W each are the realistic maximum for that edition in one workstation; three or four Max-Q cards need a chassis built for them with strong directed airflow. Our guide to GPU memory for fine-tuning explains the method.

Server. Once a model is chosen or fine-tuned, a server with RTX PRO 6000 Server Edition cards, or H200 NVL for larger training runs, serves it to the whole company next to the chat assistant; our Otter alternative comparison covers the web app staff would use.

Running the German models in production

• Voxtral Small needs vLLM 0.10.0 or later and mistral_common 1.8.1 or later; split recordings into pieces of up to 30 minutes.
• Qwen3-ASR serves through its own vLLM backend with OpenAI-compatible endpoints; timestamps come from Qwen3-ForcedAligner-0.6B, which handles up to 5 minutes of speech in 11 languages.
• Cohere Transcribe benefits from a voice activity detector or noise gate in front of it, its release blog notes, to reduce hallucinations on quiet noise.
• Comparisons. The Handy figures are Q8_0 runs in transcribe.cpp, so a BF16 deployment in vLLM can score differently; compute WER with one normaliser for all models.
• Data. Recordings of colleagues used for tests or fine-tuning are personal data, so your data-protection officer confirms on which basis you may use them. Keep test recordings and reference transcripts on the machine, and check that no tool sends audio out, as in our guide to speaker diarisation on-premise, which also covers speaker labels.

Send us the languages, dialects and hours of audio you expect, and we suggest the best solution for your tasks and budget.

FAQ