speechbrain: [Bug]: Speed of the Transcription and its Accuracy

Describe the bug

I am using EncoderDecoderASR to run inference on my local model. It takes over 10 minutes to transcribe a single wav file of length 8-9 seconds. Moreover, it gives me transcription that is completely irrelevant to the input wav file.

Transcription: ‘WHAT ARE WE GOING TO DO WHAT ARE WE GOING TO DO WHAT WILL WE DO WHAT WILL WE DO WHAT WILL WE DO WHAT WILL WE DO WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WILL WE DO GOD KNOWS WHAT WOULD WE DO GOD KNOWS WHAT WOULD WE DO GOD KNOWS WHAT WOULD WE DO GOD KNOWS WHAT WOULD WE DO GOD KNOWS WHAT WOULD WE DO GOD WOULD WE DO GOD KNOWS WHAT WOULD WE DO GOD ANY GOD KNOWS GOD KNOWS WHAT GOD KNOWS WHAT GOD IS GOD IS GOD IS GOD KNOWS WHAT GOD IS GOD AND GOD IS GOD IS GOD’

Expected behaviour

EncoderDecoderASR should give me the correct transcription, i.e, ‘PLEASE TAKE THE SHAPE OF A LONG ROUND ARCH …’ and should take less than 10 seconds to give me that output. The input is the example.wav file that is comes with SpeechBrain by default.

To Reproduce

from speechbrain.pretrained import EncoderDecoderASR
asr_model = EncoderDecoderASR.from_hparams(source="save/CKPT+2022-10-21+09-43-33+00", hparams_file="/home/axy327/speechbrain/recipes/LibriSpeech/ASR/transformer/hparams/conformer_inf.yaml", savedir="pretrained_models")
asr_model.transcribe_file("/home/axy327/speechbrain/example.wav")

The corresponding yaml code. This is very similar to the LibriSpeech conformer yaml code. I made changes to suit my environment and to suit EncoderDecoderASR() after looking at the huggingface yaml.

# ############################################################################
# Model: E2E ASR with Transformer
# Encoder: Transformer Encoder
# Decoder: Transformer Decoder + (CTC/ATT joint) beamsearch + TransformerLM
# Tokens: unigram
# losses: CTC + KLdiv (Label Smoothing loss)
# Training: Librispeech 960h
# Authors:  Jianyuan Zhong, Titouan Parcollet 2021
# ############################################################################

# Feature parameters
sample_rate: 16000
n_fft: 400
n_mels: 80

####################### Model parameters ###########################
# Transformer
d_model: 512
nhead: 4
num_encoder_layers: 8
num_decoder_layers: 6
d_ffn: 2048
transformer_dropout: 0.1
activation: !name:torch.nn.GELU
output_neurons: 5000
vocab_size: 5000

# Outputs
blank_index: 0
label_smoothing: 0.1
pad_index: 0
bos_index: 1
eos_index: 2

# Decoding parameters
min_decode_ratio: 0.0
max_decode_ratio: 1.0
valid_search_interval: 10
valid_beam_size: 10
test_beam_size: 66
lm_weight: 0.60
ctc_weight_decode: 0.40

############################## models ################################

CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
    input_shape: (8, 10, 80)
    num_blocks: 3
    num_layers_per_block: 1
    out_channels: (64, 64, 64)
    kernel_sizes: (5, 5, 1)
    strides: (2, 2, 1)
    residuals: (False, False, True)

Transformer: !new:speechbrain.lobes.models.transformer.TransformerASR.TransformerASR
    input_size: 1280
    tgt_vocab: !ref <output_neurons>
    d_model: !ref <d_model>
    nhead: !ref <nhead>
    num_encoder_layers: !ref <num_encoder_layers>
    num_decoder_layers: !ref <num_decoder_layers>
    d_ffn: !ref <d_ffn>
    dropout: !ref <transformer_dropout>
    activation: !ref <activation>
    encoder_module: conformer
    attention_type: RelPosMHAXL
    normalize_before: True
    causal: False

ctc_lin: !new:speechbrain.nnet.linear.Linear
    input_size: !ref <d_model>
    n_neurons: !ref <output_neurons>

seq_lin: !new:speechbrain.nnet.linear.Linear
    input_size: !ref <d_model>
    n_neurons: !ref <output_neurons>

decoder: !new:speechbrain.decoders.S2STransformerBeamSearch
    modules: [!ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]
    bos_index: !ref <bos_index>
    eos_index: !ref <eos_index>
    blank_index: !ref <blank_index>
    min_decode_ratio: !ref <min_decode_ratio>
    max_decode_ratio: !ref <max_decode_ratio>
    beam_size: !ref <test_beam_size>
    ctc_weight: !ref <ctc_weight_decode>
    lm_weight: !ref <lm_weight>
    lm_modules: !ref <lm_model>
    temperature: 1.15
    temperature_lm: 1.15
    using_eos_threshold: False
    length_normalization: True

log_softmax: !new:torch.nn.LogSoftmax
    dim: -1

normalizer: !new:speechbrain.processing.features.InputNormalization
    norm_type: global

compute_features: !new:speechbrain.lobes.features.Fbank
    sample_rate: !ref <sample_rate>
    n_fft: !ref <n_fft>
    n_mels: !ref <n_mels>

# This is the Transformer LM that is used according to the Huggingface repository
# Visit the HuggingFace model corresponding to the pretrained_lm_tokenizer_path
# For more details about the model!
# NB: It has to match the pre-trained TransformerLM!!
lm_model: !new:speechbrain.lobes.models.transformer.TransformerLM.TransformerLM
    vocab: 5000
    d_model: 768
    nhead: 12
    num_encoder_layers: 12
    num_decoder_layers: 0
    d_ffn: 3072
    dropout: 0.0
    activation: !name:torch.nn.GELU
    normalize_before: False

tokenizer: !new:sentencepiece.SentencePieceProcessor

Tencoder: !new:speechbrain.lobes.models.transformer.TransformerASR.EncoderWrapper
    transformer: !ref <Transformer>

encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
    input_shape: [null, null, !ref <n_mels>]
    compute_features: !ref <compute_features>
    normalize: !ref <normalizer>
    cnn: !ref <CNN>
    transformer_encoder: !ref <Tencoder>

# Models
asr_model: !new:torch.nn.ModuleList
    - [!ref <CNN>, !ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]

modules:
   compute_features: !ref <compute_features>
   normalizer: !ref <normalizer>
   pre_transformer: !ref <CNN>
   transformer: !ref <Transformer>
   asr_model: !ref <asr_model>
   lm_model: !ref <lm_model>
   encoder: !ref <encoder>
   decoder: !ref <decoder>

# The pretrainer allows a mapping between pretrained files and instances that
# are declared in the yaml.
pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
   loadables:
      normalizer: !ref <normalizer>
      asr: !ref <asr_model>
      lm: !ref <lm_model>
      tokenizer: !ref <tokenizer>

Versions

I am in the default SpeechBrain branch, i.e., develop. git hash 64c8fa951bdade7e28c81b37452331c61af7b548

Relevant log output

Did not run into any warnings or errors.

Additional context

I have trained the model using only train-clean-100 split with the conformer.yaml hparams file in the LibriSpeech recipe and used pretrained LM as in the ASR directory inside LibriSpeech/transformer. The model converged giving me WER of 8.23 on test-clean split (understandable given that I didn’t train using the complete LibriSpeech data).

Thank you for providing this toolkit to the community!