| --- |
| license: apache-2.0 |
| datasets: |
| - GTM-UVigo/FalAI |
| - juanjucm/FLEURS-SpeechT-GL-EN |
| - juanjucm/OpenHQ-SpeechT-GL-EN |
| - proxectonos/Nos_Transcrispeech-GL |
| - proxectonos/Nos_RG-Podcast-GL |
| language: |
| - gl |
| metrics: |
| - wer |
| - cer |
| base_model: |
| - facebook/w2v-bert-2.0 |
| pipeline_tag: automatic-speech-recognition |
| --- |
| # w2v-BERT 2.0 (Galician Fine-Tuned, CTC) |
|
|
| This model is a **fine-tuned version of `facebook/w2v-bert-2.0`** for **automatic speech recognition (ASR) in Galician (gl)**, trained using a **CTC objective**. |
|
|
| The model is optimised for Galician speech and evaluated across multiple domains, including read speech, broadcast-style audio and conversational content. |
|
|
| ## Training Data |
|
|
| The model was trained on a **combined Galician ASR dataset** built from several public and curated corpora. |
| All audio was normalised to **16 kHz**, and all transcripts were standardised to a homogeneous `text` format. |
|
|
| ### Datasets Included |
|
|
| - **Common Voice v23 (Galician)** |
| - **OpenSLR Speech Translation GL-EN** (Galician side) |
| - **FLEURS GL-EN** (Galician side) |
| - **FalAI** (20% of validated split) |
| - **Transcrispeech (Galician)** |
| - **RG-Podcast (Galician)** |
|
|
| These datasets cover clean read speech, semi-spontaneous speech, and more challenging acoustic conditions. |
|
|
| ### Dataset Preparation |
|
|
| - Audio resampled to **16 kHz** |
| - Removal of empty, corrupt or invalid audio |
| - Minimum audio duration: **1 second** |
| - Text normalisation: |
| - Lowercasing |
| - Unicode normalisation |
| - Removal of punctuation |
| - Removal of empty transcripts |
|
|
| ## Tokenization and Vocabulary |
|
|
| A **character-level CTC vocabulary** was constructed specifically for Galician. |
|
|
| - Supported characters: |
| `abcdefghijklmnopqrstuvwxyzáéíóúñç` |
| - Word boundaries represented using the `|` token |
| - Special tokens: |
| - `[UNK]` |
| - `[PAD]` |
|
|
| The final vocabulary is stored in `vocab.json`. |
|
|
| ## Training Procedure |
|
|
| Fine-tuning was performed using the 🤗 Transformers `Trainer` with a **CTC loss**. |
|
|
| - **Base model**: facebook/w2v-bert-2.0 |
| - **Architecture**: Wav2Vec2BertForCTC |
| - **Adapters enabled**: Yes |
|
|
| ### Training Configuration |
|
|
| - **Effective batch size**: 16 |
| - **Per-device batch size**: 8 |
| - **Gradient accumulation steps**: 2 |
| - **Learning rate**: 5e-6 |
| - **Training epochs**: 5 |
| - **Warmup ratio**: 0.1 |
| - **Precision**: FP16 |
| - **Gradient checkpointing**: Enabled |
| - **Max gradient norm**: 1.0 |
| - **Evaluation & checkpointing**: Every 2000 steps |
| - **Checkpoint limit**: 2 |
|
|
| Audio features were extracted using `SeamlessM4TFeatureExtractor`, and text was tokenized with a custom `Wav2Vec2CTCTokenizer`. |
|
|
| ## Evaluation Results |
|
|
| Evaluation was performed on held-out splits for each corpus and on a combined test set. |
| Metrics are reported as **WER** (Word Error Rate) and **CER** (Character Error Rate). |
|
|
| ### Fine-Tuned Model Results |
|
|
| **Per-corpus results** |
|
|
| | Corpus | N | WER | CER | |
| |-----------------|-------|--------|--------| |
| | FalAI | 4776 | 0.0445 | 0.0099 | |
| | CommonVoice | 14563 | 0.0628 | 0.0124 | |
| | OpenSLR | 282 | 0.1340 | 0.0406 | |
| | FLEURS | 212 | 0.1330 | 0.0447 | |
| | Transcrispeech | 1710 | 0.1410 | 0.0481 | |
| | RG-Podcast | 2015 | 0.1692 | 0.0654 | |
|
|
| **Combined test set** |
|
|
| | Dataset | N | WER | CER | |
| |---------|-------|--------|--------| |
| | TOTAL | 23558 | 0.1163 | 0.0383 | |
|
|
| ## Comparison with Whisper |
|
|
| WER comparison against Whisper-based models evaluated on the same datasets: |
|
|
| | Corpus | w2v-BERT WER | Whisper WER | |
| |-----------------|--------------|-------------| |
| | FalAI | 0.0445 | 0.0097 | |
| | CommonVoice | 0.0628 | 0.0688 | |
| | OpenSLR | 0.1340 | 0.0808 | |
| | FLEURS | 0.1330 | 0.1980 | |
| | Transcrispeech | 0.1410 | 0.2097 | |
| | RG-Podcast | 0.1692 | — | |
|
|
| ## Intended Use and Limitations |
|
|
| This model is intended for **Galician ASR research and transcription pipelines**, particularly in CTC-based or streaming-friendly setups. |
|
|
| Performance may degrade on highly spontaneous speech or extremely noisy audio. |
| The model is **monolingual (Galician-only)** and not intended for multilingual ASR or speech translation. |
|
|
| ## 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{proxectenos2026w2v-bert-2.0-gl, |
| author = {{Proxecto Nós}}, |
| title = {{w2v-BERT 2.0} (Galician Fine-Tuned, CTC)}, |
| year = {2026}, |
| publisher = {Hugging Face}, |
| howpublished = {\url{https://huggingface.co/proxectonos/w2v-bert-2.0-gl/}}, |
| } |
| ``` |
