---
license: apache-2.0
datasets:
- proxectonos/Nos_Parlaspeech-GL
- juanjucm/OpenHQ-SpeechT-GL-EN
- juanjucm/FLEURS-SpeechT-GL-EN
- mozilla-foundation/common_voice_22_0
language:
- gl
base_model:
- nvidia/stt_es_conformer_ctc_large
library_name: nemo
tags:
- automatic-speech-recognition
- speech
- audio
- CTC
- Conformer
- Transformer
- pytorch
- NeMo
- forced-alignment
- hf-asr-leaderboard
- Riva
metrics:
- wer
model-index:
- name: stt-gl-conformer-large-ctc-v1.0
  results:
  - task:
      type: Automatic Speech Recognition
      name: automatic-speech-recognition
    dataset:
      name: Common Voice 22 (Galician)
      type: mozilla/common_voice_22_gl
      config: gl
      split: test
      args:
        language: gl
    metrics:
    - name: Test WER
      type: wer
      value: 8.67
  - task:
      type: Automatic Speech Recognition
      name: automatic-speech-recognition
    dataset:
      name: Nos ParlAspeech (GL)
      type: proxectonos/Nos_Parlaspeech-GL
      config: gl
      split: test
      args:
        language: gl
    metrics:
    - name: Test WER
      type: wer
      value: 8.67
  - task:
      type: Automatic Speech Recognition
      name: automatic-speech-recognition
    dataset:
      name: OpenHQ-SpeechT (GL-EN, GL subset)
      type: juanjucm/OpenHQ-SpeechT-GL-EN
      config: gl
      split: test
      args:
        language: gl
    metrics:
    - name: Test WER
      type: wer
      value: 8.67
  - task:
      type: Automatic Speech Recognition
      name: automatic-speech-recognition
    dataset:
      name: FLEURS (GL-EN, GL subset)
      type: juanjucm/FLEURS-SpeechT-GL-EN
      config: gl
      split: test
      args:
        language: gl
    metrics:
    - name: Test WER
      type: wer
      value: 8.67
---

# Conformer-CTC Large for Galician (v1.0)

This model performs automatic speech recognition (ASR) for **Galician**.  
It is a fine-tuned version of NVIDIA's Conformer-CTC Large model, adapted to Galician using public speech datasets.

- Input: WAV mono 16 kHz
- Output: Galician transcription (lowercase)
- Architecture: Conformer-CTC (non-autoregressive)


## Installation

```bash
pip install --upgrade pip
pip install "nemo_toolkit[asr]" torch torchaudio --index-url https://download.pytorch.org/whl/cu121
```

## Quick Usage (ASR)

```python
import nemo.collections.asr as nemo_asr

MODEL_ID = "your-username/stt-gl-conformer-large-ctc-v1.0"
asr = nemo_asr.models.EncDecCTCModelBPE.from_pretrained(MODEL_ID)

preds = asr.transcribe(["/path/to/audio.wav"])  # list of wav paths
print(preds[0])
```

## Forced Alignment (word-level timestamps)

### 1) Clone and install NeMo

```bash
git clone https://github.com/NVIDIA/NeMo
cd NeMo
pip install -r requirements/requirements_asr.txt
pip install -e .
```

### 2) Export the model to `.nemo` (if needed)

```python
import nemo.collections.asr as nemo_asr
m = nemo_asr.models.EncDecCTCModelBPE.from_pretrained("your-username/stt-gl-conformer-large-ctc-v1.0")
m.save_to("stt-gl-conformer-large-ctc-v1.0.nemo")
```

### 3) Create a minimal manifest `manifest.jsonl`

```json
{"audio_filepath": "/path/audio.wav", "text": "reference transcription in galician", "duration": 3.2}
```

### 4) Run the aligner

```bash
python tools/nemo_forced_aligner/align.py \
  model_path="stt-gl-conformer-large-ctc-v1.0.nemo" \
  manifest_filepath="manifest.jsonl" \
  output_dir="align_out" \
  batch_size=1 \
  transcribe_device="cuda" \
  viterbi_device="cuda" \
  align_using_pred_text=false \
  use_local_attention=true \
  save_output_file_formats='["ctm"]'
```

Outputs are saved as `align_out/*.ctm`.


## Training Data

| Dataset | Source |
|---|---|
| Common Voice v22 GL | Mozilla (downloaded directly) |
| Nos ParlAspeech GL | `proxectonos/Nos_Parlaspeech-GL` (Zenodo) |
| OpenHQ-SpeechT GL-EN | `juanjucm/OpenHQ-SpeechT-GL-EN` |
| FLEURS GL-EN | `juanjucm/FLEURS-SpeechT-GL-EN` |

## Model Details

- **Base model:** NVIDIA Conformer-CTC Large  
- **Framework:** NeMo Toolkit  
- **Tokenizer:** SentencePiece BPE packaged inside the `.nemo`

## Forced Alignment Performance (Word-Level)

The model was also evaluated as a **word-level forced aligner** on the *Nos ParlAspeech (GL)* corpus.  
The following metrics measure the timing accuracy of predicted word boundaries relative to ground-truth annotations:

| Metric | Value |
|-------|------:|
| Mean Word Boundary Error | **47.0 ms** |
| Standard Deviation | **55.5 ms** |
| P90 (90th percentile error) | **90.0 ms** |
| Word Boundary Precision ≤ 50 ms | **64.6%** |
| Word Boundary Precision ≤ 100 ms | **92.8%** |

This indicates that a majority of predicted **word start/end timestamps** fall within 50–100 ms of the manually annotated reference boundaries, making the model suitable for segmentation, subtitle alignment, and prosody/linguistic analysis workflows.

## Limitations

Performance may degrade on highly technical domains, noisy audio, or speakers not represented in training data.

## Contact information

For further information, send an email to proxecto.nos@usc.gal

## Licensing information

[Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0)

## Acknowledgements

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project Desarrollo de Modelos ALIA. (Esta publicación del proyecto Desarrollo de Modelos ALIA está financiada por el Ministerio para la Transformación Digital y de la Función Pública y por el Plan de Recuperación, Transformación y Resiliencia – Financiado por la Unión Europea – NextGenerationEU).

Thanks also to Balidea for the technical development of this model.

## Citation
```bib
@misc{proxectenos2026stt_gl_conformer_ctc_large_v1.0,
  author       = {{Proxecto Nós}},
  title        = {{Conformer-CTC Large for Galician} (v1.0)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/proxectonos/stt_gl_conformer_ctc_large_v1.0}},
}
```