A JAX/Flax repository for combining a pre-trained speech encoder model (e.g. Wav2Vec2, HuBERT, WavLM) with a pre-trained text decoder model (e.g. GPT2, Bart) to yield a Speech Sequence-to-Sequence (Seq2Seq) model for automatic speech recognition.

The script run_flax_speech_recognition_seq2seq.py can be used to fine-tune a Speech Seq2Seq model on one of the official speech recognition datasets or a custom dataset. It makes use of the pmap JAX operator to provide data parallelism accross GPU/TPU devices.

The modelling files are based very heavily on those from Hugging Face Transformers 🤗. This is a standalone repository to enable rapid prototyping and involvement with the community. The final modelling files and training script will be merged into Transformers 🤗 to be used with the rest of the open-source library. The final system weights will be made publicly available at huggingface.co 🚀

Figure 1: Speech-encoder text-decoder style Seq2Seq model.

To instantiate a Wav2Vec2-2-Bart model with the FlaxSpeechEncoderDecoderModel framework, run the following Python script inside the cloned repo:

from transformers import AutoFeatureExtractor, AutoTokenizer
from models.modeling_flax_speech_encoder_decoder import FlaxSpeechEncoderDecoderModel
import numpy as np

# checkpoints to leverage
encoder_id = "facebook/wav2vec2-large-lv60"
decoder_id = "facebook/bart-large"

model = FlaxSpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
    encoder_id, decoder_id, encoder_add_adapter=True, decoder_from_pt=True)

model.config.decoder_start_token_id = model.config.decoder.bos_token_id
model.config.pad_token_id = model.config.decoder.pad_token_id
model.config.eos_token_id = model.config.decoder.eos_token_id
model.config.use_cache = False
model.config.processor_class = "Wav2Vec2Processor"

# check if generation works
out = model.generate(np.ones((1, 2000)))

model.save_pretrained("./")

feature_extractor = AutoFeatureExtractor.from_pretrained(encoder_id)
feature_extractor.save_pretrained("./")
tokenizer = AutoTokenizer.from_pretrained(decoder_id)
tokenizer.save_pretrained("./")

To train the model on Librispeech ASR, run the template bash script run_seq2seq_dummy.sh.

#!/usr/bin/env bash
python run_flax_speech_recognition_seq2seq.py \
        --dataset_name="librispeech_asr" \
        --model_name_or_path="openai/whisper-small" \
        --dataset_config_name="clean" \
        --train_split_name="train.100" \
        --eval_split_name="validation" \
        --test_split_name="test" \
        --text_column_name="text" \
        --id_column_name="id" \
        --output_dir="./flax-whisper-ft-librispeech-clean" \
        --wandb_project="librispeech_clean" \
        --wandb_name="flax-whisper-ft-librispeech-clean" \
        --per_device_train_batch_size="8" \
        --per_device_eval_batch_size="2" \
        --learning_rate="1e-4" \
        --warmup_steps="500" \
        --logging_steps="25" \
        --max_steps="50000" \
        --eval_steps="10000" \
        --save_steps="10000" \
        --generation_max_length="200" \
        --generation_num_beams="5" \
        --generation_length_penalty="1.2" \
        --hidden_dropout="0.2" \
        --activation_dropout="0.2" \
        --feat_proj_dropout="0.2" \
        --overwrite_output_dir \
        --gradient_checkpointing \
        --freeze_feature_encoder \
        --predict_with_generate \
        --do_lower_case \
        --do_eval \
        --do_train \
        --do_predict \
        --push_to_hub \
        --use_auth_token

Control the precision through the --precision arg:

• Full precision (weights and optimiser in fp32): --precision=full
• Half-mixed (weights in bf16, optimiser in fp32 ): --precision=half_mixed
• Full-mixed (weights and optimiser in bf16): --precision=full_mixed