Heroico Asr Conformer Es
ESPnet2 ASR model trained on the West Point HEROICO corpus (LDC2006S37 / OpenSLR SLR39, Apache 2.0).
Model source
- Kind
- Other
- File size
- 164 MB
- Version
- vRevision 0b7cb79
- Updated
- Nov 6, 2026
- Downloads
- 0
- Source
- Hugging Face
Source description
HEROICO Spanish ASR — Conformer (text-disjoint split)
ESPnet2 ASR model trained on the West Point HEROICO corpus (LDC2006S37 / OpenSLR SLR39, Apache 2.0).
- Language: Spanish (es)
- Architecture: Conformer encoder / Transformer decoder, hybrid CTC/attention, character tokenization, no LM
- Recipe:
egs2/heroico/asr1(espnet/espnet#6491), text-disjoint train/eval split - Dev: 24.4% WER / 8.1% CER (spontaneous answers, unseen speakers)
- Test: 56.3% WER / 20.6% CER — the test set is dominated by out-of-domain conditions by design; per-subset:
| test subset | utts | WER |
|---|---|---|
| answers (spontaneous, novel speakers) | 711 | 18.9% |
| recordings, clean (held-out text) | 129 | 54.3% |
| recordings, mislabeled band (corpus label noise) | 55 | 98.4% |
| usma native (held-out text + channel) | 1647 | 57.0% |
| usma non-native (held-out text + channel) | 2028 | 71.3% |
The test set looks hard because it is: 80% of it is USMA speech — a different recording site, mostly non-native speakers — reading exactly the sentences the corrected split holds out of training. See the recipe README's "Interpreting the test WER" section (espnet/espnet#6491) for the full analysis.
Note: this model replaces an earlier upload trained on a split with
~89% train/test text overlap, which scored an inflated 15.8% test WER via
text memorization (see the PR discussion). The earlier zip was also missing
exp/asr_stats_raw_es_char/train/feats_stats.npz (global-MVN stats), so it
could not be loaded as packed; this upload includes it.
Usage
With the zip (heroico_asr_conformer_v2.zip):
from espnet_model_zoo.downloader import ModelDownloader
from espnet2.bin.asr_inference import Speech2Text
d = ModelDownloader()
speech2text = Speech2Text(**d.download_and_unpack(
"https://huggingface.co/PranjulGupta/heroico-asr-conformer-es/resolve/main/heroico_asr_conformer_v2.zip"))
Or directly from the repo (unpacked layout):
from espnet2.bin.asr_inference import Speech2Text
speech2text = Speech2Text.from_pretrained("PranjulGupta/heroico-asr-conformer-es")
Input: 16 kHz mono float waveform.
@misc{LDC2006S37,
Author = {Morgan, John},
Title = {West Point Heroico Spanish Speech},
Publisher = {Linguistic Data Consortium},
Year = {2006}}
Sources
1 source
Sources
1 sourceVerified Nov 6
Model artifacts
1 artifact
Model artifacts
1 artifactexp/asr_train_asr_conformer_raw_es_char/valid.acc.ave_10best.pth
pth · 164 MB · SHA-256 33c14b1be09a…3e79 · Hugging Face
Download fileSource excerpts
3 excerpts
Source excerpts
3 excerpts--- license: apache-2.0 language: es tags: - espnet - audio - automatic-speech-recognition --- # HEROICO Spanish ASR — Conformer (text-disjoint split) ESPnet2 ASR model trained on the West Point HEROICO corpus (LDC2006S37 / OpenSLR SLR39, Apache 2.0). - **Language:** Spanish (es) - **Architecture:** Conformer encoder / Transformer decoder, hybrid CTC/attention, character tokenization, no LM - **Recipe:** `egs2/heroico/asr1` (espnet/espnet#6491), text-disjoint train/eval split - **Dev: 24.4% WER / 8.1% CER** (spontaneous answers, unseen speakers) - **Test: 56.3% WER / 20.6% CER** — the test set is dominated by out-of-domain conditions by design; per-subset: | test subset | utts | WER | |---|---|---| | answers (spontaneous, novel speakers) | 711 | 18.9% | | recordings, clean (held-out text) | 129 | 54.3% | | recordings, mislabeled band (corpus label noise) | 55 | 98.4% | | usma native (held-out text + channel) | 1647 | 57.0% | | usma non-native (held-out text + channel) | 2028 | 71.3% | The test set looks hard because it is: 80% of it is USMA speech — a different recording site, mostly non-native speakers — reading exactly the sentences the corrected split holds out of training. See the recipe README's ["Interpreting the test WER"](https://github.com/espnet/espnet/tree/master/egs2/heroico/asr1#interpreting-the-test-wer) section ([espnet/espnet#6491](https://github.com/espnet/espnet/pull/6491)) for the full analysis. **Note:** this model replaces an earlier upload trained on a split with ~89% train/test text overlap, which scored an inflated 15.8% test WER via text memorization (see the PR discussion). The earlier zip was also missing `exp/asr_stats_raw_es_char/train/feats_stats.npz` (global-MVN stats), so it could not be loaded as packed; this upload includes it. ## Usage With the zip (`heroico_asr_conformer_v2.zip`): ```python from espnet_model_zoo.downloader import ModelDownloader from espnet2.bin.asr_inference import Speech2Text d = ModelDownloader() speech2text = Speech2Text(**d.download_and_unpack( "https://huggingface.co/PranjulGupta/heroico-asr-conformer-es/resolve/main/heroico_asr_conformer_v2.zip")) ``` Or directly from the repo (unpacked layout): ```python from espnet2.bin.asr_inference import Speech2Text speech2text = Speech2Text.from_pretrained("PranjulGupta/heroico-asr-conformer-es") ``` Input: 16 kHz mono float waveform. ```...
Source context: 0 downloads · 0 likes · Pipeline automatic-speech-recognition · Library espnet · Repo PranjulGupta/heroico-asr-conformer-es
Source context: 1 downloads · 0 likes · Pipeline automatic-speech-recognition · Library espnet · Repo PranjulGupta/heroico-asr-conformer-es
