gistfile1.txt

#!/usr/bin/env python3
# gist-paste url:
# gist-paste -u https://gist.github.com/jaggzh/6e3d4d97fa93aa76320ea7b389140ff0
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from datasets import load_dataset

# This is the directory created by run_speech_recognition_seq2seq.py
whdir="whisper-custom-en"

# NOTE!  For loading large datasets, like common-voice, seq2seq's hf mozilla commonvoice loader (at least for v11)
# will try to preprocess the ENTIRE SET, even if you set your splits to small %'s.
# I modified mine to:
     #if training_args.do_train:
     #    raw_datasets["train"] = load_dataset(
     #        data_args.dataset_name,
     #        data_args.dataset_config_name,
     #        #split=data_args.train_split_name,
     #  # THIS LINE HERE AND IN THE .do_eval right below this one
     #        split=f'{data_args.train_split_name}[:1%]',  # Load only the first 1%
     #        cache_dir=model_args.cache_dir,
     #        token=model_args.token,
     #        #verification_mode='all_checks',
     #    )
# AND LOWER, BEFORE prepare_dataset(), slice the dataset (or it'll still preproc everything):
# These 4 lines:
    # if data_args.max_train_samples is not None:
    #     raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
    # if data_args.max_eval_samples is not None:
    #     raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))

    # def prepare_dataset(batch):

# load model and processor
# processor = WhisperProcessor.from_pretrained("openai/whisper-large")
# model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large")
processor = WhisperProcessor.from_pretrained(whdir)
model = WhisperForConditionalGeneration.from_pretrained(whdir)
model.config.forced_decoder_ids = None

# load dummy dataset and read audio files
#ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
#ds = load_dataset(
#    "mozilla-foundation/common_voice_11_0",
#    "en",
#    #split=data_args.train_split_name,
#    split=f'train[:15%]',  # Load only the first %
#    #cache_dir=model_args.cache_dir,
#    token=True
#    #verification_mode='all_checks',
#)
# example = ds[0]["audio"]
# sample = example['array']
# sr = example['array']

import librosa
import numpy as np
aa,sr=librosa.load("/tmp/w.wav", sr=16000)
sample=aa.astype(np.float64)
# import ipdb; ipdb.set_trace(context=16); pass
input_features = processor(sample, sampling_rate=sr, return_tensors="pt").input_features 

# generate token ids
predicted_ids = model.generate(input_features)
# decode token ids to text
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=False)
print(f"Transcription: {transcription}")

transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print(f"Transcription: {transcription}")

hf-whisper.py

#!/usr/bin/env python3
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from datasets import load_dataset

# This is the directory created by run_speech_recognition_seq2seq.py
whdir="whisper-custom-en"

# NOTE!  For loading large datasets, like common-voice, seq2seq's hf mozilla commonvoice loader (at least for v11)
# will try to preprocess the ENTIRE SET, even if you set your splits to small %'s.
# I modified mine to:
     #if training_args.do_train:
     #    raw_datasets["train"] = load_dataset(
     #        data_args.dataset_name,
     #        data_args.dataset_config_name,
     #        #split=data_args.train_split_name,
     #  # THIS LINE HERE AND IN THE .do_eval right below this one
     #        split=f'{data_args.train_split_name}[:1%]',  # Load only the first 1%
     #        cache_dir=model_args.cache_dir,
     #        token=model_args.token,
     #        #verification_mode='all_checks',
     #    )
# AND LOWER, BEFORE prepare_dataset(), slice the dataset (or it'll still preproc everything):
# These 4 lines:
    # if data_args.max_train_samples is not None:
    #     raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
    # if data_args.max_eval_samples is not None:
    #     raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))

    # def prepare_dataset(batch):

# load model and processor
# processor = WhisperProcessor.from_pretrained("openai/whisper-large")
# model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large")
processor = WhisperProcessor.from_pretrained(whdir)
model = WhisperForConditionalGeneration.from_pretrained(whdir)
model.config.forced_decoder_ids = None

# load dummy dataset and read audio files
#ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
#ds = load_dataset(
#    "mozilla-foundation/common_voice_11_0",
#    "en",
#    #split=data_args.train_split_name,
#    split=f'train[:15%]',  # Load only the first %
#    #cache_dir=model_args.cache_dir,
#    token=True
#    #verification_mode='all_checks',
#)
# example = ds[0]["audio"]
# sample = example['array']
# sr = example['array']

import librosa
import numpy as np
aa,sr=librosa.load("/tmp/w.wav", sr=16000)
sample=aa.astype(np.float64)
# import ipdb; ipdb.set_trace(context=16); pass
input_features = processor(sample, sampling_rate=sr, return_tensors="pt").input_features 

# generate token ids
predicted_ids = model.generate(input_features)
# decode token ids to text
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=False)
print(f"Transcription: {transcription}")

transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print(f"Transcription: {transcription}")