language: es
datasets:
- common_voice
- ciempiess_test
- hub4ne_es_LDC98S74
- callhome_es_LDC96S35
tags:
- audio
- automatic-speech-recognition
- spanish
- xlrs-53-spanish
- ciempiess
- cimpiess-unam
license: cc-by-4.0
widget: null
model-index:
- name: wav2vec2-large-xlsr-53-spanish-ep5-944h
results:
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: Mozilla Common Voice 10.0
type: mozilla-foundation/common_voice_10_0
split: test
args:
language: es
metrics:
- name: WER
type: wer
value: <unk>
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: Mozilla Common Voice 10.0
type: mozilla-foundation/common_voice_10_0
split: validation
args:
language: es
metrics:
- name: WER
type: wer
value: <unk>
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: CIEMPIESS-TEST
type: ciempiess/ciempiess_test
split: test
args:
language: es
metrics:
- name: WER
type: wer
value: 11.17
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: 1997 Spanish Broadcast News Speech (HUB4-NE)
type: HUB4NE_LDC98S74
split: test
args:
language: es
metrics:
- name: WER
type: wer
value: 7.48
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: CALLHOME Spanish Speech (Test)
type: callhome_LDC96S35
split: test
args:
language: es
metrics:
- name: WER
type: wer
value: 39.12
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: CALLHOME Spanish Speech (Dev)
type: callhome_LDC96S35
split: validation
args:
language: es
metrics:
- name: WER
type: wer
value: 40.39
wav2vec2-large-xlsr-53-spanish-ep5-944h
The "wav2vec2-large-xlsr-53-spanish-ep5-944h" is an acoustic model suitable for Automatic Speech Recognition in Spanish. It is the result of fine-tuning the model "facebook/wav2vec2-large-xlsr-53" with around 944 hours of Spanish data gathered or developed by the CIEMPIESS-UNAM Project since 2012. Most of the data is available at the the CIEMPIESS-UNAM Project homepage http://www.ciempiess.org/. The rest can be found in public repositories such as LDC or OpenSLR
The specific list of corpora used to fine-tune the model is:
- CIEMPIESS-LIGHT (18h25m)
- CIEMPIESS-BALANCE (18h20m)
- CIEMPIESS-FEM (13h54m)
- CHM150 (1h38m)
- TEDX_SPANISH (24h29m)
- LIBRIVOX_SPANISH (73h01m)
- WIKIPEDIA_SPANISH (25h37m)
- VOXFORGE_SPANISH (49h42m)
- MOZILLA COMMON VOICE 10.0 (320h22m)
- HEROICO (16h33m)
- LATINO-40 (6h48m)
- CALLHOME_SPANISH (13h22m)
- HUB4NE_SPANISH (31h41m)
- FISHER_SPANISH (127h22m)
- Chilean Spanish speech data set (7h08m)
- Colombian Spanish speech data set (7h34m)
- Peruvian Spanish speech data set (9h13m)
- Argentinian Spanish speech data set (8h01m)
- Puerto Rico Spanish speech data set (1h00m)
- MediaSpeech Spanish (10h00m)
- DIMEX100-LIGHT (6h09m)
- DIMEX100-NIÑOS (08h09m)
- GOLEM-UNIVERSUM (00h10m)
- GLISSANDO (6h40m)
- TELE_con_CIENCIA (28h16m) Unplished Material
- UNSHAREABLE MATERIAL (118h22m) Not available for sharing
The fine-tuning process was perform during November (2022) in the servers of the Language and Voice Lab (https://lvl.ru.is/) at Reykjavík University (Iceland) by Carlos Daniel Hernández Mena.
Evaluation
import torch
from transformers import Wav2Vec2Processor
from transformers import Wav2Vec2ForCTC
#Load the processor and model.
MODEL_NAME="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-spanish-ep5-944h"
processor = Wav2Vec2Processor.from_pretrained(MODEL_NAME)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_NAME)
#Load the dataset
from datasets import load_dataset, load_metric, Audio
ds=load_dataset("ciempiess/ciempiess_test", split="test")
#Normalize the transcriptions
import re
chars_to_ignore_regex = '[\\,\\?\\.\\!\\\;\\:\\"\\“\\%\\‘\\”\\�\\)\\(\\*)]'
def remove_special_characters(batch):
batch["normalized_text"] = re.sub(chars_to_ignore_regex, '', batch["normalized_text"]).lower()
return batch
ds = ds.map(remove_special_characters)
#Downsample to 16kHz
ds = ds.cast_column("audio", Audio(sampling_rate=16_000))
#Process the dataset
def prepare_dataset(batch):
audio = batch["audio"]
#Batched output is "un-batched" to ensure mapping is correct
batch["input_values"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0]
with processor.as_target_processor():
batch["labels"] = processor(batch["normalized_text"]).input_ids
return batch
ds = ds.map(prepare_dataset, remove_columns=ds.column_names,num_proc=1)
#Define the evaluation metric
import numpy as np
wer_metric = load_metric("wer")
def compute_metrics(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id
pred_str = processor.batch_decode(pred_ids)
#We do not want to group tokens when computing the metrics
label_str = processor.batch_decode(pred.label_ids, group_tokens=False)
wer = wer_metric.compute(predictions=pred_str, references=label_str)
return {"wer": wer}
#Do the evaluation (with batch_size=1)
model = model.to(torch.device("cuda"))
def map_to_result(batch):
with torch.no_grad():
input_values = torch.tensor(batch["input_values"], device="cuda").unsqueeze(0)
logits = model(input_values).logits
pred_ids = torch.argmax(logits, dim=-1)
batch["pred_str"] = processor.batch_decode(pred_ids)[0]
batch["normalized_text"] = processor.decode(batch["labels"], group_tokens=False)
return batch
results = ds.map(map_to_result,remove_columns=ds.column_names)
#Compute the overall WER now.
print("Test WER: {:.3f}".format(wer_metric.compute(predictions=results["pred_str"], references=results["normalized_text"])))
Test Result: 0.112
BibTeX entry and citation info
When publishing results based on these models please refer to:
@misc{mena2022xlrs53spanish,
title={Acoustic Model in Spanish: wav2vec2-large-xlsr-53-spanish-ep5-944h.},
author={Hernandez Mena, Carlos Daniel},
year={2022},
url={https://huggingface.co/carlosdanielhernandezmena/wav2vec2-large-xlsr-53-spanish-ep5-944h},
}
Acknowledgements
The author wants to thank to the social service program "Desarrollo de Tecnologías del Habla" at the Facultad de Ingeniería (FI) of the Universidad Nacional Autónoma de México (UNAM). He also thanks to the social service students for all the hard work.
Special thanks to Jón Guðnason, head of the Language and Voice Lab for providing computational power to make this model possible. The author also thanks to the "Language Technology Programme for Icelandic 2019-2023" which is managed and coordinated by Almannarómur, and it is funded by the Icelandic Ministry of Education, Science and Culture.