README.md

---
language:
- da
license: other
license_name: ekko-composite-model-terms
license_link: https://huggingface.co/RyeAI/ekko-v1-tiny/blob/main/MODEL_LICENSE.md
thumbnail: https://huggingface.co/RyeAI/ekko-v1-tiny/resolve/e37e698805c538e447c25062ab13fa862462ee9e/cover.png
library_name: nemo
pipeline_tag: automatic-speech-recognition
base_model: nvidia/parakeet-rnnt-110m-da-dk
base_model_relation: finetune
datasets:
- alexandrainst/nst-da
- alexandrainst/ftspeech
- alexandrainst/nota
- CoRal-project/coral-v3
tags:
- automatic-speech-recognition
- speech-recognition
- speech-to-text
- danish
- danish-asr
- offline
- local-inference
- cpu-inference
- nemo
- gguf
- onnx
- sherpa-onnx
- rnnt
- word-timestamps
metrics:
- wer
model-index:
- name: Ekko v1 Tiny
  results:
  - task:
      type: automatic-speech-recognition
    dataset:
      name: CoRal conversation
      type: CoRal-project/coral-v3
      config: conversation
      split: test
    metrics:
    - type: wer
      value: 24.15
  - task:
      type: automatic-speech-recognition
    dataset:
      name: CoRal read aloud
      type: CoRal-project/coral-v3
      config: read_aloud
      split: test
    metrics:
    - type: wer
      value: 11.87
  - task:
      type: automatic-speech-recognition
    dataset:
      name: FLEURS Danish
      type: google/fleurs
      config: da_dk
      split: test
    metrics:
    - type: wer
      value: 9.75
  - task:
      type: automatic-speech-recognition
    dataset:
      name: FTSpeech
      type: alexandrainst/ftspeech
      split: test_balanced
    metrics:
    - type: wer
      value: 7.62
---

![Ekko v1 Tiny — offline Danish speech recognition](cover.png)

[Leaderboard](https://huggingface.co/spaces/RyeAI/danish-asr-leaderboard) ·
[Benchmark data](https://huggingface.co/datasets/RyeAI/danish-asr-leaderboard) ·
[Evaluation harness](https://github.com/Rye-A1/danish-asr-leaderboard)

# Ekko v1 Tiny — Offline Danish Speech-to-Text with Word Timestamps

Ekko v1 Tiny is a compact, 110M-parameter Danish speech recognition model with
word timestamps. It is a FastConformer RNN-T fine-tuned from
`nvidia/parakeet-rnnt-110m-da-dk` and runs locally. Its raw output is lowercase
text; pair it with `RyeAI/ekko-pnc` for capitalization and punctuation.

**Best fit:** clear Danish speech from one speaker, delivered at a steady pace,
such as dictation or reading aloud. The published results are stronger on read
and formal speech than on spontaneous conversation. Dictation itself was not
evaluated as a separate test set.

## Benchmark

**12.37% mean WER** across 26,780 rows from five Danish test sets, evaluated
with the leaderboard harness. Scores use the published defaults: NFKC
normalization, spoken-number folding, Danish number-word normalization,
lowercase, punctuation removal, symmetric filler removal, and whitespace
collapse. Lower is better.

| Test set | Speech type | WER |
|---|---|---:|
| CoRal conversation | Spontaneous conversation | 24.15% |
| CoRal read aloud | Read speech | 11.87% |
| Common Voice Danish (CV 25.0 test subset) | Crowd-sourced read speech | 8.44% |
| FLEURS `da_dk` | Read news-style speech | 9.75% |
| FTSpeech | Parliamentary speech | 7.62% |
| **Equal-weight mean** | | **12.37%** |

The NeMo checkpoint processed the benchmark at **795.8x realtime** on one A100.
The official run saved raw per-sample outputs, and an independent offline
rescore reproduced 12.37% exactly.

### Comparison with models below 800M parameters

![Mean WER of Ekko v1 Tiny and all twelve open-weight leaderboard models below 800M parameters](comparison.png)

The chart compares Tiny's Q5_0 release score with 12 open-weight models below
800M parameters in the [Danish ASR Leaderboard](https://huggingface.co/datasets/RyeAI/danish-asr-leaderboard/tree/f7c930a30ac2f88e68421d88fc577c0cbafa8c31/results)
as of 9 October 2026. Each score is an equal-weight mean WER across the same
five Danish test sets (26,780 clips). Tiny scores 12.37%, 0.27 percentage
points lower than Svale 600M at 12.64%. Rankings may change when new models
or reruns are added.

### Native GGUF quantization

The F32, F16, Q8_0, Q6_K, Q5_K, Q5_0, Q4_K, and Q4_0 files are
scored on the same current 26,780-row matrix and normalization policy as the
NeMo result above. All 40 variant/domain outputs completed and passed exact-row
offline scoring.

| Variant | Size | Mean WER | CPU speed |
|---|---:|---:|---:|
| **F16** | 248.2 MiB | **12.315%** | 74.4x |
| F32 | 430.6 MiB | 12.318% | 71.8x |
| **Q8_0** | 162.7 MiB | **12.323%** | **74.9x** |
| Q6_K | 141.8 MiB | 12.325% | 63.0x |
| Q5_K | 129.8 MiB | 12.337% | 60.1x |
| **Q5_0** | **128.5 MiB** | **12.370%** | **69.4x** |
| Q4_K | 118.4 MiB | 12.442% | 65.4x |
| Q4_0 | 117.1 MiB | 12.467% | 72.1x |

Q5_0 remains the packaged default: it is 128.5 MiB, runs at 69.4x realtime,
and is 0.047 mean-WER points behind Q8_0. Choose Q8_0 when the extra 34.2 MiB
is acceptable and maximum quantized accuracy is preferred. F16 is the native
reference tier.

### ONNX and browser export

The model is available for sherpa-onnx and was evaluated on the same 26,780-row
matrix with mandatory non-silent peak normalization to 0.95. The browser
downloads the compact int8 export.

| Export | Size | Mean WER | CPU speed |
|---|---:|---:|---:|
| FP32 ONNX | 450.6 MiB | 12.465% | 53.4x |
| **INT8 ONNX** | **129.0 MiB** | **12.946%** | **49.2x** |

Both exports completed all five domains with zero runtime failures. INT8 is
71.4% smaller than FP32 and trades 0.481 mean-WER points for the smaller browser
download.

## Artifacts

The current NeMo, GGUF, and ONNX artifacts are pinned at Hub revision
`e37e698805c538e447c25062ab13fa862462ee9e` and verified by SHA-256.

Clean-cache packaged download and inference tests passed on macOS arm64 and
Linux x64 with native word timestamps. The Windows x64 artifact is mapped but
has not been exercised against this release.

| Format | Path | Measured mean WER |
|---|---|---:|
| NeMo | `ekko-v1-tiny.nemo` | 12.37% |
| parakeet.cpp GGUF | `gguf-parakeet.cpp/ekko-v1-tiny-{f32,f16,q8_0,q6_k,q5_k,q5_0,q4_k,q4_0}.gguf` | 12.315-12.467% |
| sherpa-onnx FP32 | `onnx-sherpa/{encoder,decoder,joiner}.onnx` | 12.465% |
| sherpa-onnx INT8 | `onnx-sherpa/{encoder,decoder,joiner}.int8.onnx` | 12.946% |

The browser loads `onnx-sherpa/runtime-config.json` with the INT8 files and
checks its SHA-256 before using its audio and model settings.

## Local Use

For pinned downloads, word timestamps, and optional PnC, see the
[Ekko STT quickstart](https://github.com/Rye-A1/ekko-danish-stt#run-locally).
If you already have `parakeet-cli` from parakeet.cpp, you can also run the
GGUF model directly:

```bash
hf download RyeAI/ekko-v1-tiny \
  gguf-parakeet.cpp/ekko-v1-tiny-q5_0.gguf \
  --local-dir ./ekko-v1-tiny
parakeet-cli transcribe \
  --model ./ekko-v1-tiny/gguf-parakeet.cpp/ekko-v1-tiny-q5_0.gguf \
  --input audio.wav --timestamps
```

NeMo usage:

```python
from nemo.collections.asr.models import EncDecRNNTBPEModel

model = EncDecRNNTBPEModel.from_pretrained("RyeAI/ekko-v1-tiny")
transcript = model.transcribe(["audio.wav"])[0]
print(transcript.text)
```

## Training and limitations

Ekko v1 Tiny trained for 50,000 steps on an 879,083-row,
1,549.6-hour Danish manifest, including 454.1 hours of quality-gated
pseudo-labelled audio.

- Danish only.
- Raw output has no punctuation or capitalization.
- The 44-token vocabulary does not contain uppercase or punctuation tokens.
- Punctuation and capitalization require a separate restoration model such as
  `RyeAI/ekko-pnc`.
- Spontaneous conversation remains the hardest benchmark domain (24.15% WER).
- Noisy audio and overlapping speakers were not evaluated separately.

## License

The complete model and data terms are documented in
[`MODEL_LICENSE.md`](MODEL_LICENSE.md) and the bundled `LICENSES/` files.