Although OpenAI open-sourced multi-lingual Whisper model, https://github.com/openai/whisper, achived the state-of-art results in the benchmark dataset, there are many scenaria the pretrained models donot work well. For example, the languages not covered in the pretrained model. Whisper-V3 support 100 languages. Thus, the model must be re-trained in order to support new language. For the minority languages, even they are covered in the pretrained model, the accuracy is often worse and need to collect more data to adapt the pretrained model in order to reduce word error rate. In the post, huggingface implemented whisper version is used to fine-tune the Chinese language (It is just for testing training code functionality rather than training a product version model. No computing resource. As soon as resources (computing and data) ready, SOTA model can be trained). You can refer Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers . The following figure shows the logic flow for model training.
- Get annotated training data, i.e. a set of speech-text pair samples. In the demo code, I get the data from the common-voice (Chinese)
- Download the pretrained whisper based model
- Transform speech wave data to Mel-spectrum features, which are further feeded into transformer encoder
- For Whisper, speech must be 16K sample rate. If not, nee re-sampling
- The size of Mel band is 80 or 128 (for whisper large)
- Transformer encoder-decoder is trained to learn text-audio cross attension and audio self-attention. The predicted next token, P(next-token|text, audio) (the probability is calculated in the decoder), together with ground-truth text to calculate cross-entropy loss. Then gradients are computed and model parameters are updated.
The training code and word-error-rate (WER) evaluation code is in the following:
"""
Test codes for fine-tuning Whisper speech-to-text, i.e. speech recognition
"""
from datasets import load_dataset, DatasetDict, Audio
from transformers import WhisperFeatureExtractor, WhisperTokenizer, WhisperProcessor, WhisperForConditionalGeneration
import torch
from dataclasses import dataclass
from typing import Any, Dict, List, Union
import evaluate
from transformers import Seq2SeqTrainingArguments
from transformers import Seq2SeqTrainer
from evaluate import load
@dataclass
class DataCollatorSpeechSeq2SeqWithPadding:
processor: Any
def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
# split inputs and labels since they have to be of different lengths and need different padding methods
# first treat the audio inputs by simply returning torch tensors
input_features = [{"input_features": feature["input_features"]} for feature in features]
batch = self.processor.feature_extractor.pad(input_features, return_tensors="pt")
# get the tokenized label sequences
label_features = [{"input_ids": feature["labels"]} for feature in features]
# pad the labels to max length
labels_batch = self.processor.tokenizer.pad(label_features, return_tensors="pt")
# replace padding with -100 to ignore loss correctly
labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
# if bos token is appended in previous tokenization step,
# cut bos token here as it's append later anyways
if (labels[:, 0] == self.processor.tokenizer.bos_token_id).all().cpu().item():
labels = labels[:, 1:]
batch["labels"] = labels
return batch
def compute_metrics(pred):
pred_ids = pred.predictions
label_ids = pred.label_ids
# replace -100 with the pad_token_id
label_ids[label_ids == -100] = tokenizer.pad_token_id
# we do not want to group tokens when computing the metrics
pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
label_str = tokenizer.batch_decode(label_ids, skip_special_tokens=True)
wer = 100 * metric.compute(predictions=pred_str, references=label_str)
return {"wer": wer}
def prepare_dataset(batch):
# load and resample audio data from 48 to 16kHz
audio = batch["audio"]
# print(f"""** {batch}""")
# compute log-Mel input features from input audio array
batch["input_features"] = feature_extractor(audio["array"], sampling_rate=audio["sampling_rate"]).input_features[0]
# encode target text to label ids
batch["labels"] = tokenizer(batch["sentence"]).input_ids
# print(f"""** {batch["labels"]}""")
return batch
#Step-1: Define model structure & initialization, feature extractor, text tokenizer
model_base_default = "openai/whisper-small"
language = "zh"
save_dir = "whisper-small-zh-me"
max_steps = 500
feature_extractor = WhisperFeatureExtractor.from_pretrained(model_base_default, language=language)
tokenizer = WhisperTokenizer.from_pretrained(model_base_default, task="transcribe", language=language)
processor = WhisperProcessor.from_pretrained(model_base_default, task="transcribe", language=language)
#Step-2: Prepare train/dev/test data
common_voice1 = DatasetDict()
common_voice1["train"] = load_dataset("mozilla-foundation/common_voice_11_0", "zh-CN", split="train", use_auth_token=False).select(range(1000))
common_voice1["test"] = load_dataset("mozilla-foundation/common_voice_11_0", "zh-CN", split="test", use_auth_token=False).select(range(50))
common_voice = common_voice1.remove_columns(["accent", "age", "client_id", "down_votes", "gender", "locale", "path", "segment", "up_votes"])
common_voice = common_voice.cast_column("audio", Audio(sampling_rate=16000))
common_voice = common_voice.map(prepare_dataset, remove_columns=common_voice.column_names["train"], num_proc=1)
data_collator = DataCollatorSpeechSeq2SeqWithPadding(processor=processor)
metric = evaluate.load("wer")
#
model = WhisperForConditionalGeneration.from_pretrained(model_base_default)
model.generation_config.language = language
model.config.forced_decoder_ids = None
model.config.suppress_tokens = []
training_args = Seq2SeqTrainingArguments(
output_dir=save_dir, #"./whisper-small-zh-me", # change to a repo name of your choice
per_device_train_batch_size=16,
gradient_accumulation_steps=1, # increase by 2x for every 2x decrease in batch size
learning_rate=1e-5,
warmup_steps=100, #500,
max_steps=max_steps, #4000,
gradient_checkpointing=True,
fp16=True,
evaluation_strategy="steps",
per_device_eval_batch_size=8,
predict_with_generate=True,
generation_max_length=225,
save_steps=500, #500, #1000,
eval_steps=500, #500, #1000,
logging_steps=25,
report_to=["tensorboard"],
load_best_model_at_end=True,
metric_for_best_model="wer",
greater_is_better=False,
push_to_hub=False, #True,
)
trainer = Seq2SeqTrainer(
args=training_args,
model=model,
train_dataset=common_voice["train"],
eval_dataset=common_voice["test"],
data_collator=data_collator,
compute_metrics=compute_metrics,
tokenizer=processor.feature_extractor,
)
trainer.train()
#evaluate the model trained
#evaluation
# from datasets import load_dataset
# from transformers import WhisperForConditionalGeneration, WhisperProcessor
# import torch
# from evaluate import load
# from datasets import Audio
# librispeech_test_clean = load_dataset("librispeech_asr", "clean", split="test")
model_pth = f"{save_dir}/checkpoint-{max_steps}" #(
model_token_pth = "openai/whisper-small"
# model_pth = "openai/whisper-large"
is_local = True
processor = WhisperProcessor.from_pretrained(model_token_pth)#"openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained(model_pth, local_files_only=is_local).to("cuda")
def map_to_pred(batch):
audio = batch["audio"]
input_features = processor(audio["array"], sampling_rate=audio["sampling_rate"], return_tensors="pt").input_features
batch["reference"] = processor.tokenizer._normalize(batch['sentence'])
with torch.no_grad():
predicted_ids = model.generate(input_features.to("cuda"))[0]
transcription = processor.decode(predicted_ids)
batch["prediction"] = processor.tokenizer._normalize(transcription)
return batch
common_voice2 = common_voice1["test"].cast_column("audio", Audio(sampling_rate=16000))
result = common_voice2.map(map_to_pred)
wer = load("wer")
print(100 * wer.compute(references=result["reference"], predictions=result["prediction"]))
The above code is tested in 4070.
-
Previous
Document summarization and question-answering on document -LLM + LangChain -
Next
How to read annual report
