FunASR/egs/alimeeting/sa-asr/asr_local_m2met_2023_infer.sh

#!/usr/bin/env bash

# Set bash to 'debug' mode, it will exit on :
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
set -e
set -u
set -o pipefail

log() {
    local fname=${BASH_SOURCE[1]##*/}
    echo -e "$(date '+%Y-%m-%dT%H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
}
min() {
  local a b
  a=$1
  for b in "$@"; do
      if [ "${b}" -le "${a}" ]; then
          a="${b}"
      fi
  done
  echo "${a}"
}
SECONDS=0

# General configuration
stage=1              # Processes starts from the specified stage.
stop_stage=10000     # Processes is stopped at the specified stage.
skip_data_prep=false # Skip data preparation stages.
skip_train=false     # Skip training stages.
skip_eval=false      # Skip decoding and evaluation stages.
skip_upload=true     # Skip packing and uploading stages.
ngpu=1               # The number of gpus ("0" uses cpu, otherwise use gpu).
num_nodes=1          # The number of nodes.
nj=16                # The number of parallel jobs.
inference_nj=16      # The number of parallel jobs in decoding.
gpu_inference=false  # Whether to perform gpu decoding.
njob_infer=4
dumpdir=dump2         # Directory to dump features.
expdir=exp           # Directory to save experiments.
python=python3       # Specify python to execute espnet commands.
device=0

# Data preparation related
local_data_opts= # The options given to local/data.sh.

# Speed perturbation related
speed_perturb_factors=  # perturbation factors, e.g. "0.9 1.0 1.1" (separated by space).

# Feature extraction related
feats_type=raw       # Feature type (raw or fbank_pitch).
audio_format=flac    # Audio format: wav, flac, wav.ark, flac.ark  (only in feats_type=raw).
fs=16000             # Sampling rate.
min_wav_duration=0.1 # Minimum duration in second.
max_wav_duration=20  # Maximum duration in second.

# Tokenization related
token_type=bpe      # Tokenization type (char or bpe).
nbpe=30             # The number of BPE vocabulary.
bpemode=unigram     # Mode of BPE (unigram or bpe).
oov="<unk>"         # Out of vocabulary symbol.
blank="<blank>"     # CTC blank symbol
sos_eos="<sos/eos>" # sos and eos symbole
bpe_input_sentence_size=100000000 # Size of input sentence for BPE.
bpe_nlsyms=         # non-linguistic symbols list, separated by a comma, for BPE
bpe_char_cover=1.0  # character coverage when modeling BPE

# Language model related
use_lm=true       # Use language model for ASR decoding.
lm_tag=           # Suffix to the result dir for language model training.
lm_exp=           # Specify the direcotry path for LM experiment.
                  # If this option is specified, lm_tag is ignored.
lm_stats_dir=     # Specify the direcotry path for LM statistics.
lm_config=        # Config for language model training.
lm_args=          # Arguments for language model training, e.g., "--max_epoch 10".
                  # Note that it will overwrite args in lm config.
use_word_lm=false # Whether to use word language model.
num_splits_lm=1   # Number of splitting for lm corpus.
# shellcheck disable=SC2034
word_vocab_size=10000 # Size of word vocabulary.

# ASR model related
asr_tag=       # Suffix to the result dir for asr model training.
asr_exp=       # Specify the direcotry path for ASR experiment.
               # If this option is specified, asr_tag is ignored.
sa_asr_exp=
asr_stats_dir= # Specify the direcotry path for ASR statistics.
asr_config=    # Config for asr model training.
sa_asr_config=
asr_args=      # Arguments for asr model training, e.g., "--max_epoch 10".
               # Note that it will overwrite args in asr config.
feats_normalize=global_mvn # Normalizaton layer type.
num_splits_asr=1           # Number of splitting for lm corpus.

# Decoding related
inference_tag=    # Suffix to the result dir for decoding.
inference_config= # Config for decoding.
inference_args=   # Arguments for decoding, e.g., "--lm_weight 0.1".
                  # Note that it will overwrite args in inference config.
sa_asr_inference_tag=
sa_asr_inference_args=

inference_lm=valid.loss.ave.pb        # Language modle path for decoding.
inference_asr_model=valid.acc.ave.pb  # ASR model path for decoding.
                                      # e.g.
                                      # inference_asr_model=train.loss.best.pth
                                      # inference_asr_model=3epoch.pth
                                      # inference_asr_model=valid.acc.best.pth
                                      # inference_asr_model=valid.loss.ave.pth
inference_sa_asr_model=valid.acc_spk.ave.pb
download_model= # Download a model from Model Zoo and use it for decoding.

# [Task dependent] Set the datadir name created by local/data.sh
train_set=       # Name of training set.
valid_set=       # Name of validation set used for monitoring/tuning network training.
test_sets=       # Names of test sets. Multiple items (e.g., both dev and eval sets) can be specified.
bpe_train_text=  # Text file path of bpe training set.
lm_train_text=   # Text file path of language model training set.
lm_dev_text=     # Text file path of language model development set.
lm_test_text=    # Text file path of language model evaluation set.
nlsyms_txt=none  # Non-linguistic symbol list if existing.
cleaner=none     # Text cleaner.
g2p=none         # g2p method (needed if token_type=phn).
lang=zh      # The language type of corpus.
score_opts=                # The options given to sclite scoring
local_score_opts=          # The options given to local/score.sh.

help_message=$(cat << EOF
Usage: $0 --train-set "<train_set_name>" --valid-set "<valid_set_name>" --test_sets "<test_set_names>"

Options:
    # General configuration
    --stage          # Processes starts from the specified stage (default="${stage}").
    --stop_stage     # Processes is stopped at the specified stage (default="${stop_stage}").
    --skip_data_prep # Skip data preparation stages (default="${skip_data_prep}").
    --skip_train     # Skip training stages (default="${skip_train}").
    --skip_eval      # Skip decoding and evaluation stages (default="${skip_eval}").
    --skip_upload    # Skip packing and uploading stages (default="${skip_upload}").
    --ngpu           # The number of gpus ("0" uses cpu, otherwise use gpu, default="${ngpu}").
    --num_nodes      # The number of nodes (default="${num_nodes}").
    --nj             # The number of parallel jobs (default="${nj}").
    --inference_nj   # The number of parallel jobs in decoding (default="${inference_nj}").
    --gpu_inference  # Whether to perform gpu decoding (default="${gpu_inference}").
    --dumpdir        # Directory to dump features (default="${dumpdir}").
    --expdir         # Directory to save experiments (default="${expdir}").
    --python         # Specify python to execute espnet commands (default="${python}").
    --device         # Which GPUs are use for local training (defalut="${device}").

    # Data preparation related
    --local_data_opts # The options given to local/data.sh (default="${local_data_opts}").

    # Speed perturbation related
    --speed_perturb_factors # speed perturbation factors, e.g. "0.9 1.0 1.1" (separated by space, default="${speed_perturb_factors}").

    # Feature extraction related
    --feats_type       # Feature type (raw, fbank_pitch or extracted, default="${feats_type}").
    --audio_format     # Audio format: wav, flac, wav.ark, flac.ark  (only in feats_type=raw, default="${audio_format}").
    --fs               # Sampling rate (default="${fs}").
    --min_wav_duration # Minimum duration in second (default="${min_wav_duration}").
    --max_wav_duration # Maximum duration in second (default="${max_wav_duration}").

    # Tokenization related
    --token_type              # Tokenization type (char or bpe, default="${token_type}").
    --nbpe                    # The number of BPE vocabulary (default="${nbpe}").
    --bpemode                 # Mode of BPE (unigram or bpe, default="${bpemode}").
    --oov                     # Out of vocabulary symbol (default="${oov}").
    --blank                   # CTC blank symbol (default="${blank}").
    --sos_eos                 # sos and eos symbole (default="${sos_eos}").
    --bpe_input_sentence_size # Size of input sentence for BPE (default="${bpe_input_sentence_size}").
    --bpe_nlsyms              # Non-linguistic symbol list for sentencepiece, separated by a comma. (default="${bpe_nlsyms}").
    --bpe_char_cover          # Character coverage when modeling BPE (default="${bpe_char_cover}").

    # Language model related
    --lm_tag          # Suffix to the result dir for language model training (default="${lm_tag}").
    --lm_exp          # Specify the direcotry path for LM experiment.
                      # If this option is specified, lm_tag is ignored (default="${lm_exp}").
    --lm_stats_dir    # Specify the direcotry path for LM statistics (default="${lm_stats_dir}").
    --lm_config       # Config for language model training (default="${lm_config}").
    --lm_args         # Arguments for language model training (default="${lm_args}").
                      # e.g., --lm_args "--max_epoch 10"
                      # Note that it will overwrite args in lm config.
    --use_word_lm     # Whether to use word language model (default="${use_word_lm}").
    --word_vocab_size # Size of word vocabulary (default="${word_vocab_size}").
    --num_splits_lm   # Number of splitting for lm corpus (default="${num_splits_lm}").

    # ASR model related
    --asr_tag          # Suffix to the result dir for asr model training (default="${asr_tag}").
    --asr_exp          # Specify the direcotry path for ASR experiment.
                       # If this option is specified, asr_tag is ignored (default="${asr_exp}").
    --asr_stats_dir    # Specify the direcotry path for ASR statistics (default="${asr_stats_dir}").
    --asr_config       # Config for asr model training (default="${asr_config}").
    --asr_args         # Arguments for asr model training (default="${asr_args}").
                       # e.g., --asr_args "--max_epoch 10"
                       # Note that it will overwrite args in asr config.
    --feats_normalize  # Normalizaton layer type (default="${feats_normalize}").
    --num_splits_asr   # Number of splitting for lm corpus  (default="${num_splits_asr}").

    # Decoding related
    --inference_tag       # Suffix to the result dir for decoding (default="${inference_tag}").
    --inference_config    # Config for decoding (default="${inference_config}").
    --inference_args      # Arguments for decoding (default="${inference_args}").
                          # e.g., --inference_args "--lm_weight 0.1"
                          # Note that it will overwrite args in inference config.
    --inference_lm        # Language modle path for decoding (default="${inference_lm}").
    --inference_asr_model # ASR model path for decoding (default="${inference_asr_model}").
    --download_model      # Download a model from Model Zoo and use it for decoding (default="${download_model}").

    # [Task dependent] Set the datadir name created by local/data.sh
    --train_set     # Name of training set (required).
    --valid_set     # Name of validation set used for monitoring/tuning network training (required).
    --test_sets     # Names of test sets.
                    # Multiple items (e.g., both dev and eval sets) can be specified (required).
    --bpe_train_text # Text file path of bpe training set.
    --lm_train_text  # Text file path of language model training set.
    --lm_dev_text   # Text file path of language model development set (default="${lm_dev_text}").
    --lm_test_text  # Text file path of language model evaluation set (default="${lm_test_text}").
    --nlsyms_txt    # Non-linguistic symbol list if existing (default="${nlsyms_txt}").
    --cleaner       # Text cleaner (default="${cleaner}").
    --g2p           # g2p method (default="${g2p}").
    --lang          # The language type of corpus (default=${lang}).
    --score_opts             # The options given to sclite scoring (default="{score_opts}").
    --local_score_opts       # The options given to local/score.sh (default="{local_score_opts}").
EOF
)

log "$0 $*"
# Save command line args for logging (they will be lost after utils/parse_options.sh)
run_args=$(python -m funasr.utils.cli_utils $0 "$@")
. utils/parse_options.sh

if [ $# -ne 0 ]; then
    log "${help_message}"
    log "Error: No positional arguments are required."
    exit 2
fi

. ./path.sh


# Check required arguments
[ -z "${train_set}" ] && { log "${help_message}"; log "Error: --train_set is required"; exit 2; };
[ -z "${valid_set}" ] && { log "${help_message}"; log "Error: --valid_set is required"; exit 2; };
[ -z "${test_sets}" ] && { log "${help_message}"; log "Error: --test_sets is required"; exit 2; };

# Check feature type
if [ "${feats_type}" = raw ]; then
    data_feats=${dumpdir}/raw
elif [ "${feats_type}" = fbank_pitch ]; then
    data_feats=${dumpdir}/fbank_pitch
elif [ "${feats_type}" = fbank ]; then
    data_feats=${dumpdir}/fbank
elif [ "${feats_type}" == extracted ]; then
    data_feats=${dumpdir}/extracted
else
    log "${help_message}"
    log "Error: not supported: --feats_type ${feats_type}"
    exit 2
fi

# Use the same text as ASR for bpe training if not specified.
[ -z "${bpe_train_text}" ] && bpe_train_text="${data_feats}/${train_set}/text"
# Use the same text as ASR for lm training if not specified.
[ -z "${lm_train_text}" ] && lm_train_text="${data_feats}/${train_set}/text"
# Use the same text as ASR for lm training if not specified.
[ -z "${lm_dev_text}" ] && lm_dev_text="${data_feats}/${valid_set}/text"
# Use the text of the 1st evaldir if lm_test is not specified
[ -z "${lm_test_text}" ] && lm_test_text="${data_feats}/${test_sets%% *}/text"

# Check tokenization type
if [ "${lang}" != noinfo ]; then
    token_listdir=data/${lang}_token_list
else
    token_listdir=data/token_list
fi
bpedir="${token_listdir}/bpe_${bpemode}${nbpe}"
bpeprefix="${bpedir}"/bpe
bpemodel="${bpeprefix}".model
bpetoken_list="${bpedir}"/tokens.txt
chartoken_list="${token_listdir}"/char/tokens.txt
# NOTE: keep for future development.
# shellcheck disable=SC2034
wordtoken_list="${token_listdir}"/word/tokens.txt

if [ "${token_type}" = bpe ]; then
    token_list="${bpetoken_list}"
elif [ "${token_type}" = char ]; then
    token_list="${chartoken_list}"
    bpemodel=none
elif [ "${token_type}" = word ]; then
    token_list="${wordtoken_list}"
    bpemodel=none
else
    log "Error: not supported --token_type '${token_type}'"
    exit 2
fi
if ${use_word_lm}; then
    log "Error: Word LM is not supported yet"
    exit 2

    lm_token_list="${wordtoken_list}"
    lm_token_type=word
else
    lm_token_list="${token_list}"
    lm_token_type="${token_type}"
fi


# Set tag for naming of model directory
if [ -z "${asr_tag}" ]; then
    if [ -n "${asr_config}" ]; then
        asr_tag="$(basename "${asr_config}" .yaml)_${feats_type}"
    else
        asr_tag="train_${feats_type}"
    fi
    if [ "${lang}" != noinfo ]; then
        asr_tag+="_${lang}_${token_type}"
    else
        asr_tag+="_${token_type}"
    fi
    if [ "${token_type}" = bpe ]; then
        asr_tag+="${nbpe}"
    fi
    # Add overwritten arg's info
    if [ -n "${asr_args}" ]; then
        asr_tag+="$(echo "${asr_args}" | sed -e "s/--/\_/g" -e "s/[ |=/]//g")"
    fi
    if [ -n "${speed_perturb_factors}" ]; then
        asr_tag+="_sp"
    fi
fi
if [ -z "${lm_tag}" ]; then
    if [ -n "${lm_config}" ]; then
        lm_tag="$(basename "${lm_config}" .yaml)"
    else
        lm_tag="train"
    fi
    if [ "${lang}" != noinfo ]; then
        lm_tag+="_${lang}_${lm_token_type}"
    else
        lm_tag+="_${lm_token_type}"
    fi
    if [ "${lm_token_type}" = bpe ]; then
        lm_tag+="${nbpe}"
    fi
    # Add overwritten arg's info
    if [ -n "${lm_args}" ]; then
        lm_tag+="$(echo "${lm_args}" | sed -e "s/--/\_/g" -e "s/[ |=/]//g")"
    fi
fi

# The directory used for collect-stats mode
if [ -z "${asr_stats_dir}" ]; then
    if [ "${lang}" != noinfo ]; then
        asr_stats_dir="${expdir}/asr_stats_${feats_type}_${lang}_${token_type}"
    else
        asr_stats_dir="${expdir}/asr_stats_${feats_type}_${token_type}"
    fi
    if [ "${token_type}" = bpe ]; then
        asr_stats_dir+="${nbpe}"
    fi
    if [ -n "${speed_perturb_factors}" ]; then
        asr_stats_dir+="_sp"
    fi
fi
if [ -z "${lm_stats_dir}" ]; then
    if [ "${lang}" != noinfo ]; then
        lm_stats_dir="${expdir}/lm_stats_${lang}_${lm_token_type}"
    else
        lm_stats_dir="${expdir}/lm_stats_${lm_token_type}"
    fi
    if [ "${lm_token_type}" = bpe ]; then
        lm_stats_dir+="${nbpe}"
    fi
fi
# The directory used for training commands
if [ -z "${asr_exp}" ]; then
    asr_exp="${expdir}/asr_${asr_tag}"
fi
if [ -z "${lm_exp}" ]; then
    lm_exp="${expdir}/lm_${lm_tag}"
fi


if [ -z "${inference_tag}" ]; then
    if [ -n "${inference_config}" ]; then
        inference_tag="$(basename "${inference_config}" .yaml)"
    else
        inference_tag=inference
    fi
    # Add overwritten arg's info
    if [ -n "${inference_args}" ]; then
        inference_tag+="$(echo "${inference_args}" | sed -e "s/--/\_/g" -e "s/[ |=]//g")"
    fi
    if "${use_lm}"; then
        inference_tag+="_lm_$(basename "${lm_exp}")_$(echo "${inference_lm}" | sed -e "s/\//_/g" -e "s/\.[^.]*$//g")"
    fi
    inference_tag+="_asr_model_$(echo "${inference_asr_model}" | sed -e "s/\//_/g" -e "s/\.[^.]*$//g")"
fi

if [ -z "${sa_asr_inference_tag}" ]; then
    if [ -n "${inference_config}" ]; then
        sa_asr_inference_tag="$(basename "${inference_config}" .yaml)"
    else
        sa_asr_inference_tag=sa_asr_inference
    fi
    # Add overwritten arg's info
    if [ -n "${sa_asr_inference_args}" ]; then
        sa_asr_inference_tag+="$(echo "${sa_asr_inference_args}" | sed -e "s/--/\_/g" -e "s/[ |=]//g")"
    fi
    if "${use_lm}"; then
        sa_asr_inference_tag+="_lm_$(basename "${lm_exp}")_$(echo "${inference_lm}" | sed -e "s/\//_/g" -e "s/\.[^.]*$//g")"
    fi
    sa_asr_inference_tag+="_asr_model_$(echo "${inference_sa_asr_model}" | sed -e "s/\//_/g" -e "s/\.[^.]*$//g")"
fi

train_cmd="run.pl"
cuda_cmd="run.pl"
decode_cmd="run.pl"

# ========================== Main stages start from here. ==========================

if ! "${skip_data_prep}"; then

    if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
        if [ "${feats_type}" = raw ]; then
            log "Stage 1: Format wav.scp: data/ -> ${data_feats}"

            # ====== Recreating "wav.scp" ======
            # Kaldi-wav.scp, which can describe the file path with unix-pipe, like "cat /some/path |",
            # shouldn't be used in training process.
            # "format_wav_scp.sh" dumps such pipe-style-wav to real audio file
            # and it can also change the audio-format and sampling rate.
            # If nothing is need, then format_wav_scp.sh does nothing:
            # i.e. the input file format and rate is same as the output.

            for dset in "${test_sets}" ; do
            
                _suf=""

                local/copy_data_dir.sh --validate_opts --non-print data/"${dset}" "${data_feats}${_suf}/${dset}"
                
                rm -f ${data_feats}${_suf}/${dset}/{segments,wav.scp,reco2file_and_channel,reco2dur}
                _opts=
                if [ -e data/"${dset}"/segments ]; then
                    # "segments" is used for splitting wav files which are written in "wav".scp
                    # into utterances. The file format of segments:
                    #   <segment_id> <record_id> <start_time> <end_time>
                    #   "e.g. call-861225-A-0050-0065 call-861225-A 5.0 6.5"
                    # Where the time is written in seconds.
                    _opts+="--segments data/${dset}/segments "
                fi
                # shellcheck disable=SC2086
                local/format_wav_scp.sh --nj "${nj}" --cmd "${train_cmd}" \
                    --audio-format "${audio_format}" --fs "${fs}" ${_opts} \
                    "data/${dset}/wav.scp" "${data_feats}${_suf}/${dset}"

                echo "${feats_type}" > "${data_feats}${_suf}/${dset}/feats_type"
            done

        else
            log "Error: not supported: --feats_type ${feats_type}"
            exit 2
        fi
    fi

    if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
        log "Stage 2: Generate speaker profile by spectral-cluster"
        mkdir -p "profile_log"
        for dset in "${test_sets}"; do
            # generate cluster_profile with spectral-cluster directly (for infering and without oracle information)
            python local/gen_cluster_profile_infer.py "${data_feats}/${dset}" "data/${dset}" 0.996 0.815 &> "profile_log/gen_cluster_profile_infer_${dset}.log"
            log "Successfully generate cluster profile for ${dset} (${data_feats}/${dset}/cluster_profile_infer.scp)"
            done
    fi

else
    log "Skip the stages for data preparation"
fi


# ========================== Data preparation is done here. ==========================

if ! "${skip_eval}"; then

    if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
        log "Stage 3: Decoding SA-ASR (cluster profile): training_dir=${sa_asr_exp}"

        if ${gpu_inference}; then
            _cmd="${cuda_cmd}"
            inference_nj=$[${ngpu}*${njob_infer}]
            _ngpu=1

        else
            _cmd="${decode_cmd}"
            inference_nj=$njob_infer
            _ngpu=0
        fi

        _opts=
        if [ -n "${inference_config}" ]; then
            _opts+="--config ${inference_config} "
        fi
        if "${use_lm}"; then
            if "${use_word_lm}"; then
                _opts+="--word_lm_train_config ${lm_exp}/config.yaml "
                _opts+="--word_lm_file ${lm_exp}/${inference_lm} "
            else
                _opts+="--lm_train_config ${lm_exp}/config.yaml "
                _opts+="--lm_file ${lm_exp}/${inference_lm} "
            fi
        fi

        # 2. Generate run.sh
        log "Generate '${sa_asr_exp}/${sa_asr_inference_tag}.cluster/run.sh'. You can resume the process from stage 17 using this script"
        mkdir -p "${sa_asr_exp}/${sa_asr_inference_tag}.cluster"; echo "${run_args} --stage 17 \"\$@\"; exit \$?" > "${sa_asr_exp}/${sa_asr_inference_tag}.cluster/run.sh"; chmod +x "${sa_asr_exp}/${sa_asr_inference_tag}.cluster/run.sh"

        for dset in ${test_sets}; do
            _data="${data_feats}/${dset}"
            _dir="${sa_asr_exp}/${sa_asr_inference_tag}.cluster/${dset}"
            _logdir="${_dir}/logdir"
            mkdir -p "${_logdir}"

            _feats_type="$(<${_data}/feats_type)"
            if [ "${_feats_type}" = raw ]; then
                _scp=wav.scp
                if [[ "${audio_format}" == *ark* ]]; then
                    _type=kaldi_ark
                else
                    _type=sound
                fi
            else
                _scp=feats.scp
                _type=kaldi_ark
            fi

            # 1. Split the key file
            key_file=${_data}/${_scp}
            split_scps=""
            _nj=$(min "${inference_nj}" "$(<${key_file} wc -l)")
            for n in $(seq "${_nj}"); do
                split_scps+=" ${_logdir}/keys.${n}.scp"
            done
            # shellcheck disable=SC2086
            utils/split_scp.pl "${key_file}" ${split_scps}

            # 2. Submit decoding jobs
            log "Decoding started... log: '${_logdir}/sa_asr_inference.*.log'"
            # shellcheck disable=SC2086
            ${_cmd} --gpu "${_ngpu}" --max-jobs-run "${_nj}" JOB=1:"${_nj}" "${_logdir}"/asr_inference.JOB.log \
                python -m funasr.bin.asr_inference_launch \
                    --batch_size 1 \
                    --mc True   \
                    --nbest 1   \
                    --ngpu "${_ngpu}" \
                    --njob ${njob_infer} \
                    --gpuid_list ${device} \
                    --data_path_and_name_and_type "${_data}/${_scp},speech,${_type}" \
                    --data_path_and_name_and_type "${_data}/cluster_profile_infer.scp,profile,npy" \
                    --key_file "${_logdir}"/keys.JOB.scp \
                    --allow_variable_data_keys true \
                    --asr_train_config "${sa_asr_exp}"/config.yaml \
                    --asr_model_file "${sa_asr_exp}"/"${inference_sa_asr_model}" \
                    --output_dir "${_logdir}"/output.JOB \
                    --mode sa_asr \
                    ${_opts}

            # 3. Concatenates the output files from each jobs
            for f in token token_int score text text_id; do
                for i in $(seq "${_nj}"); do
                    cat "${_logdir}/output.${i}/1best_recog/${f}"
                done | LC_ALL=C sort -k1 >"${_dir}/${f}"
            done
        done
    fi

    if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
        log "Stage 4: Generate SA-ASR results (cluster profile)"

        for dset in ${test_sets}; do
            _dir="${sa_asr_exp}/${sa_asr_inference_tag}.cluster/${dset}"

            python local/process_text_spk_merge.py ${_dir}
        done

    fi

else
    log "Skip the evaluation stages"
fi


log "Successfully finished. [elapsed=${SECONDS}s]"