Instructions to use moravianai/CzechSpeech with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use moravianai/CzechSpeech with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="moravianai/CzechSpeech")# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq processor = AutoProcessor.from_pretrained("moravianai/CzechSpeech") model = AutoModelForSpeechSeq2Seq.from_pretrained("moravianai/CzechSpeech", device_map="auto") - Notebooks
- Google Colab
- Kaggle
CzechSpeech: Czech speech to text on the device
CzechSpeech is a speech-to-text model (Whisper large-v3-turbo) fine-tuned on Czech speech by Moravian AI. It transcribes parliamentary speech, read speech and speech in noise more accurately than available open models of comparable size. Quantised to 4 bits it fits in 452 MiB and runs offline on a laptop or on an ordinary phone.
Dual licence: free for non-commercial use under CC BY-NC 4.0; a commercial licence is available from Moravian AI: info@moravian.ai.
Word error rate: lower is better
Word error rate (WER) on the same segments with the same text normalisation and greedy decoding for every model, laptop with an RTX 4050. Lower is better.
| model | parameters | parliament | FLEURS | noise |
|---|---|---|---|---|
| CzechSpeech | 0.81 B | 1.8 % | 7.3 % | 6.5 % |
| NVIDIA Canary-1b-v2 | 0.98 B | 7.5 % | 8.4 % | 22.4 % |
| Whisper large-v3, fine-tuned on Czech Common Voice 13 | 1.55 B | 8.0 % | 9.6 % | 29.1 % |
| Whisper large-v3 | 1.55 B | 8.1 % | 10.2 % | 27.8 % |
| Whisper large-v3-turbo (the starting point) | 0.81 B | 9.5 % | 11.4 % | 29.5 % |
| NVIDIA Parakeet-TDT-0.6b-v3 | 0.6 B | 10.3 % | 11.9 % | 25.8 % |
Numbers and legal terms: higher is better
Share transcribed correctly on the parliamentary set: numeric entities (numbers, dates, amounts) and legal terms from our list. Higher is better. Measured for the models below.
| model | numbers correct | legal terms correct |
|---|---|---|
| CzechSpeech | 99.3 % | 99.5 % |
| NVIDIA Canary-1b-v2 | 94.8 % | 94.2 % |
| Whisper large-v3-turbo (the starting point) | 96.5 % | 96.3 % |
| NVIDIA Parakeet-TDT-0.6b-v3 | 94.6 % | 92.1 % |
Test sets
- Parliament: the test split of ParCzech4Speech, 729 segments (106 min): random segments, segments with numbers and segments with legal terms. The speakers of the test and validation splits were removed from training.
- FLEURS: the Czech test set, 723 sentences (147 min).
- Noise: the parliamentary segments with reverberation and crowd babble at 10 dB signal-to-noise ratio.
- Parliamentary speech is the training domain of CzechSpeech and the FLEURS training split was in its training data too. The least favourable comparison is therefore FLEURS: 7.3 % against 8.4 % for Canary-1b-v2.
- CzechSpeech in whisper.cpp without timestamps (8-bit file); Whisper large-v3 and its Czech fine-tune in whisper.cpp; Canary-1b-v2 and Parakeet-TDT-0.6b-v3 in ONNX (fp32).
Files
| file | for | size | WER parliament / FLEURS / noise (lower is better) |
|---|---|---|---|
model.safetensors + config and tokenizer |
transformers, 16-bit | 1.5 GiB | 1.8 / 7.6 / 6.4 % (transformers, greedy) |
ggml-czechspeech-q5_0.bin |
whisper.cpp on a laptop or desktop (GPU or CPU) | 547 MiB | 1.8 / 7.4 / 6.6 % |
ggml-czechspeech-q4_0.bin |
whisper.cpp on phones and ARM CPUs | 452 MiB | 1.9 / 7.7 / 6.8 % |
The whisper.cpp files are GPTQ quantisations with Czech calibration data (512 segments from the training mix) onto the q5_0 and q4_0 grids: 5 bits lose nothing against 8 bits, 4 bits lose 0.1 to 0.3 points of WER. On ARM CPUs with the i8mm instructions q4_0 is about 4× faster than q5_0, because ggml repacks q4_0 weights for them.
| device | file | memory | speed |
|---|---|---|---|
| Pixel 9a phone, processor only, 4 threads | q4_0 | 0.83 GB RAM at the peak | being re-measured |
| Laptop with an RTX 4050 (6 GB) | q5_0 | 0.89 GB GPU memory | 0.33 s per segment (median), about 25× faster than real time |
Phone speed was measured before the encoder window got its 10-second minimum (see Short utterances below) and is being re-measured, so no figure is given here.
Usage
transformers:
import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor
processor = WhisperProcessor.from_pretrained("moravianai/CzechSpeech")
model = WhisperForConditionalGeneration.from_pretrained("moravianai/CzechSpeech", dtype=torch.float16).to("cuda")
# audio: a float32 array, 16 kHz, mono, up to 30 s
features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features.to("cuda", torch.float16)
ids = model.generate(features, language="cs", task="transcribe")
print(processor.batch_decode(ids, skip_special_tokens=True)[0])
whisper.cpp:
whisper-cli -m ggml-czechspeech-q5_0.bin -l cs -nt -f recording.wav
- Decode without timestamps (
-ntin whisper.cpp). CzechSpeech was trained without timestamps; with timestamps whisper.cpp sometimes starts the transcript with a stray word. - Short utterances: the encoder can process a window shorter than the fixed 30 seconds (
audio_ctxin whisper.cpp), which cuts the encoder work, but do not go below 512 frames (about 10 s). CzechSpeech was trained on full windows: with the window trimmed to the speech, utterances under 4 s got invented or repeated sentences in 15 to 46 % of cases. Our SDKs use ceil(samples / 320) + 32 frames, rounded up to a multiple of 16, at least 512 and at most 1500; on short utterances this lowered WER from 6.28 to 3.90 % on clean audio and from 10.82 to 9.89 % in noise, compared with the trimmed window. - Recordings longer than 30 s need to be split, as with any Whisper model.
Use in an app
Two SDKs run this model on the device, with speech detection (Silero VAD) and whisper.cpp:
- Android:
implementation("ai.moravian:czechspeech-android:0.3.0")from Maven Central; see moravian-ai/czechspeech-android. - Web:
npm install @moravianai/czechspeech(WebGPU, or WebAssembly on the CPU); see moravian-ai/czechspeech-web.
Both download this model's q4_0 file from a pinned revision, verify its SHA-256 and keep audio on the device. The SDKs have a licence of their own: free for noncommercial use (PolyForm Noncommercial 1.0.0); commercial licence: info@moravian.ai.
Training
- Base model: openai/whisper-large-v3-turbo.
- Data: ParCzech4Speech 1.0, unsegmented variant (1 336 h, CC BY 4.0), YODAS-Granary Czech (65 h, twice, CC BY 3.0), FLEURS cs_cz train (8.4 h, three times, CC BY 4.0) and VoxPopuli Czech (5.8 h, twice, CC0); 1 503 h in total, segments of 1 to 30 s, without timestamps. The speakers of the evaluation splits were removed.
- Compute: 24 968 steps on the LUMI supercomputer.
Limitations
- Most of the training data is parliamentary speech: a formal register with many numbers and legal terms. Spontaneous conversation, dialects and telephone audio were not measured.
- Transcription only: no translation, speaker separation or timestamps; no analysis of emotions or tone of voice.
- Numbers are written mostly as digits, following the ParCzech transcripts.
Enterprise edition
Need more than Czech? CzechSpeech Enterprise is our offer for companies that put speech recognition into their own products, under a commercial licence; you can test it first under a test licence.
- Czech, Slovak and English in one model (beta, available). Its Czech word error rate is the same as this model's on parliament speech (1.9 %), lower on FLEURS (7.3 % against 7.8 %) and 6.8 % against 6.7 % on noisy speech, within the margin of the measurement (4-bit files, same whisper.cpp build); Slovak FLEURS 6.6 % and English FLEURS 6.6 % WER. It does not pass every point of our Czech release gate and it drops part of the speech played through a loudspeaker; the Enterprise page lists all of it.
- A Czech language model that picks the best of several transcript variants and joins sentences split by a pause (a converted third-party model, MIT).
- A variant for meetings and calls (in evaluation). On our test sets, against this model (4-bit files, same
whisper.cpp build):
- a voice played through a loudspeaker (a remote participant, a phone on speaker), with or without noise: 9.5 % WER against 19.3 %;
- someone cutting in while another person speaks: 4.2 % of the speaker's words lost against 7.0 %;
- a loud voice from a loudspeaker under the speaker: 12.9 % WER against 17.2 %;
- two people taking turns: both written in 97 % of pairs against 94.5 %.
- Speaker recognition (who said what) with a voiceprint model fine-tuned on Czech speech: a pilot on request, after legal review. A person's voiceprint is stored only with that person's explicit consent, on the device.
- Tuned to your vocabulary and your recordings, with the accuracy measured on your own data before you decide.
- On your users' devices, like this model: audio stays on the device, with no licence server and no telemetry.
To try it, fill in the form on the CzechSpeech Enterprise page (organisation, contact e-mail, phone, acceptance of the test licence); you get access at once. For production and commercial use, a licence from us: info@moravian.ai.
Licence and attribution
CzechSpeech weights, dual licence: © 2026 Jan Rosecký (Moravian AI).
- Non-commercial use: free under CC BY-NC 4.0.
- Commercial use: a separate commercial licence from Moravian AI, write to info@moravian.ai.
Base model: Whisper large-v3-turbo, © 2022 OpenAI, MIT licence (see
NOTICE.md).Training data: ParCzech4Speech 1.0 (Stankov, Kopp et al., ÚFAL MFF UK, arXiv 2509.06675), YODAS-Granary (ESPnet, NVIDIA Granary), FLEURS (Google, arXiv 2205.12446) and VoxPopuli (Meta).
Header painting: © Jan Bébar, used with the author's permission.
Acknowledgements
We acknowledge the EuroHPC Joint Undertaking for awarding this project access to the EuroHPC supercomputer LUMI, hosted by CSC (Finland) and the LUMI consortium.
Results and details: moravian.ai/czechspeech · contact: info@moravian.ai
- Downloads last month
- 412
