|
| 1 | +#!/usr/bin/env bash |
| 2 | + |
| 3 | +set -ex |
| 4 | + |
| 5 | +log() { |
| 6 | + # This function is from espnet |
| 7 | + local fname=${BASH_SOURCE[1]##*/} |
| 8 | + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" |
| 9 | +} |
| 10 | + |
| 11 | +repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09 |
| 12 | +repo=$(basename $repo_url) |
| 13 | +log "Download pretrained model and test-data from $repo_url" |
| 14 | + |
| 15 | +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url |
| 16 | +pushd $repo |
| 17 | +git lfs pull --include "exp/cpu_jit.pt" |
| 18 | + |
| 19 | +git lfs pull --include "data/lang_bpe_500/tokens.txt" |
| 20 | +git lfs pull --include "data/lang_bpe_500/HLG.pt" |
| 21 | +git lfs pull --include "data/lang_bpe_500/words.txt" |
| 22 | + |
| 23 | +git lfs pull --include "data/lm/G_4_gram.pt" |
| 24 | +popd |
| 25 | + |
| 26 | +log "Test CTC decode (librispeech)" |
| 27 | + |
| 28 | +./build/bin/ctc_decode \ |
| 29 | + --use_gpu false \ |
| 30 | + --nn_model $repo/exp/cpu_jit.pt \ |
| 31 | + --tokens $repo/data/lang_bpe_500/tokens.txt \ |
| 32 | + $repo/test_wavs/1089-134686-0001.wav \ |
| 33 | + $repo/test_wavs/1221-135766-0001.wav \ |
| 34 | + $repo/test_wavs/1221-135766-0002.wav |
| 35 | + |
| 36 | + |
| 37 | +log "Test HLG decode (librispeech)" |
| 38 | + |
| 39 | +./build/bin/hlg_decode \ |
| 40 | + --use_gpu false \ |
| 41 | + --nn_model $repo/exp/cpu_jit.pt \ |
| 42 | + --hlg $repo/data/lang_bpe_500/HLG.pt \ |
| 43 | + --word_table $repo/data/lang_bpe_500/words.txt \ |
| 44 | + $repo/test_wavs/1089-134686-0001.wav \ |
| 45 | + $repo/test_wavs/1221-135766-0001.wav \ |
| 46 | + $repo/test_wavs/1221-135766-0002.wav |
| 47 | + |
| 48 | +if [ $(uname) == "Darwin" ]; then |
| 49 | + # GitHub only provides 7 GB RAM for Linux/Windows |
| 50 | + # It has 14 GB RAM for macOS. This test requires a lot of RAM. |
| 51 | + log "Test n-gram LM rescore (librispeech)" |
| 52 | + ./build/bin/ngram_lm_rescore \ |
| 53 | + --use_gpu false \ |
| 54 | + --nn_model $repo/exp/cpu_jit.pt \ |
| 55 | + --hlg $repo/data/lang_bpe_500/HLG.pt \ |
| 56 | + --g $repo/data/lm/G_4_gram.pt \ |
| 57 | + --ngram_lm_scale 1.0 \ |
| 58 | + --word_table $repo/data/lang_bpe_500/words.txt \ |
| 59 | + $repo/test_wavs/1089-134686-0001.wav \ |
| 60 | + $repo/test_wavs/1221-135766-0001.wav \ |
| 61 | + $repo/test_wavs/1221-135766-0002.wav |
| 62 | + |
| 63 | + log "Test n-gram LM rescore + attention rescore (librispeech)" |
| 64 | + ./build/bin/attention_rescore \ |
| 65 | + --use_gpu false \ |
| 66 | + --nn_model $repo/exp/cpu_jit.pt \ |
| 67 | + --hlg $repo/data/lang_bpe_500/HLG.pt \ |
| 68 | + --g $repo/data/lm/G_4_gram.pt \ |
| 69 | + --ngram_lm_scale 1.0 \ |
| 70 | + --attention_scale 1.0 \ |
| 71 | + --num_paths 100 \ |
| 72 | + --nbest_scale 0.5 \ |
| 73 | + --word_table $repo/data/lang_bpe_500/words.txt \ |
| 74 | + --sos_id 1 \ |
| 75 | + --eos_id 1 \ |
| 76 | + $repo/test_wavs/1089-134686-0001.wav \ |
| 77 | + $repo/test_wavs/1221-135766-0001.wav \ |
| 78 | + $repo/test_wavs/1221-135766-0002.wav |
| 79 | +fi |
| 80 | + |
| 81 | +log "Streaming CTC decoding" |
| 82 | + |
| 83 | +./build/bin/online_decode \ |
| 84 | + --use_ctc_decoding true \ |
| 85 | + --jit_pt $repo/exp/cpu_jit.pt \ |
| 86 | + --tokens $repo/data/lang_bpe_500/tokens.txt \ |
| 87 | + $repo/test_wavs/1089-134686-0001.wav \ |
| 88 | + $repo/test_wavs/1221-135766-0001.wav \ |
| 89 | + $repo/test_wavs/1221-135766-0002.wav |
| 90 | + |
| 91 | +log "Streaming HLG decoding" |
| 92 | + |
| 93 | +./build/bin/online_decode \ |
| 94 | + --use_ctc_decoding false \ |
| 95 | + --jit_pt $repo/exp/cpu_jit.pt \ |
| 96 | + --hlg $repo/data/lang_bpe_500/HLG.pt \ |
| 97 | + --word_table $repo/data/lang_bpe_500/words.txt \ |
| 98 | + $repo/test_wavs/1089-134686-0001.wav \ |
| 99 | + $repo/test_wavs/1221-135766-0001.wav \ |
| 100 | + $repo/test_wavs/1221-135766-0002.wav |
| 101 | + |
| 102 | +rm -rf repo |
| 103 | + |
| 104 | +# Now for RNN-T |
| 105 | + |
| 106 | +repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13 |
| 107 | +repo=$(basename $repo_url) |
| 108 | +log "Download pretrained model and test-data from $repo_url" |
| 109 | + |
| 110 | +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url |
| 111 | +pushd $repo |
| 112 | +git lfs pull --include "exp/cpu_jit.pt" |
| 113 | +git lfs pull --include "data/lang_bpe_500/LG.pt" |
| 114 | +popd |
| 115 | + |
| 116 | +log "Test RNN-T decoding" |
| 117 | + |
| 118 | +./build/bin/pruned_stateless_transducer \ |
| 119 | + --use-gpu=false \ |
| 120 | + --nn-model=$repo/exp/cpu_jit.pt \ |
| 121 | + --tokens=$repo/data/lang_bpe_500/tokens.txt \ |
| 122 | + $repo/test_wavs/1089-134686-0001.wav \ |
| 123 | + $repo/test_wavs/1221-135766-0001.wav \ |
| 124 | + $repo/test_wavs/1221-135766-0002.wav |
| 125 | + |
| 126 | +./build/bin/rnnt_demo \ |
| 127 | + --use_lg false \ |
| 128 | + --jit_pt $repo/exp/cpu_jit.pt \ |
| 129 | + --tokens $repo/data/lang_bpe_500/tokens.txt \ |
| 130 | + $repo/test_wavs/1089-134686-0001.wav \ |
| 131 | + $repo/test_wavs/1221-135766-0001.wav \ |
| 132 | + $repo/test_wavs/1221-135766-0002.wav |
| 133 | + |
| 134 | +./build/bin/rnnt_demo \ |
| 135 | + --use_lg true \ |
| 136 | + --jit_pt $repo/exp/cpu_jit.pt \ |
| 137 | + --lg $repo/data/lang_bpe_500/LG.pt \ |
| 138 | + --word_table $repo/data/lang_bpe_500/words.txt \ |
| 139 | + --beam 8 \ |
| 140 | + --max_contexts 8 \ |
| 141 | + --max_states 64 \ |
| 142 | + $repo/test_wavs/1089-134686-0001.wav \ |
| 143 | + $repo/test_wavs/1221-135766-0001.wav \ |
| 144 | + $repo/test_wavs/1221-135766-0002.wav |
0 commit comments