--- tags: - espnet - audio - automatic-speech-summarization language: en datasets: - how2 license: cc-by-4.0 --- ## ESPnet2 ASR model ### `espnet/roshansh_how2_asr_raw_ft_sum_valid.acc` This model was trained by roshansh-cmu using how2 recipe in [espnet](https://github.com/espnet/espnet/). ### Demo: How to use in ESPnet2 ```bash cd espnet git checkout e6f42a9783a5d9eba0687c19417f933e890722d7 pip install -e . cd egs2/how2/sum1 ./run.sh --skip_data_prep false --skip_train true --download_model espnet/roshansh_how2_asr_raw_ft_sum_valid.acc ``` # RESULTS ## Environments - date: `Mon Feb 7 15:24:21 EST 2022` - python version: `3.8.12 (default, Oct 12 2021, 13:49:34) [GCC 7.5.0]` - espnet version: `espnet 0.10.6a1` - pytorch version: `pytorch 1.10.1` - Git hash: `04561cdf3b6c3bc1d51edb04c93b953759ef551d` - Commit date: `Mon Feb 7 09:06:12 2022 -0500` ## asr_raw_ft_sum |dataset|Snt|Wrd|ROUGE-1|ROUGE-2|ROUGE-L|METEOR|BERTScore| |---|---|---|---|---|---|---|---| |decode_sum_asr_model_valid.acc.best/dev5_test_sum|2127|69795|60.72|44.7|56.1|29.36|91.53| ## ASR config
expand ``` config: conf/train_asr_conformer_vid_lf.yaml print_config: false log_level: INFO dry_run: false iterator_type: sequence output_dir: exp/asr_raw_ft_sum ngpu: 1 seed: 0 num_workers: 1 num_att_plot: 3 dist_backend: nccl dist_init_method: env:// dist_world_size: 8 dist_rank: 0 local_rank: 0 dist_master_addr: localhost dist_master_port: 45875 dist_launcher: null multiprocessing_distributed: true unused_parameters: true sharded_ddp: false cudnn_enabled: true cudnn_benchmark: false cudnn_deterministic: true collect_stats: false write_collected_feats: false max_epoch: 100 patience: 10 val_scheduler_criterion: - valid - loss early_stopping_criterion: - valid - loss - min best_model_criterion: - - valid - acc - max keep_nbest_models: 10 grad_clip: 5.0 grad_clip_type: 2.0 grad_noise: false accum_grad: 10 no_forward_run: false resume: true train_dtype: float32 use_amp: false log_interval: 5000 use_tensorboard: true use_wandb: false wandb_project: null wandb_id: null wandb_entity: null wandb_name: null wandb_model_log_interval: -1 detect_anomaly: false pretrain_path: null init_param: - exp/asr_raw_utt_conformer/valid.acc.ave_10best.pth:::ctc ignore_init_mismatch: false freeze_param: [] num_iters_per_epoch: null batch_size: 20 valid_batch_size: null batch_bins: 60000000 valid_batch_bins: null train_shape_file: - exp/asr_stats_raw_vid_sum/train/speech_shape - exp/asr_stats_raw_vid_sum/train/text_shape.bpe valid_shape_file: - exp/asr_stats_raw_vid_sum/valid/speech_shape - exp/asr_stats_raw_vid_sum/valid/text_shape.bpe batch_type: length valid_batch_type: null fold_length: - 80000 - 150 sort_in_batch: descending sort_batch: descending multiple_iterator: false chunk_length: 500 chunk_shift_ratio: 0.5 num_cache_chunks: 1024 train_data_path_and_name_and_type: - - dump/raw/tr_2000h_sum_trim/wav.scp - speech - sound - - dump/raw/tr_2000h_sum_trim/text - text - text valid_data_path_and_name_and_type: - - dump/raw/cv05_sum_trim/wav.scp - speech - sound - - dump/raw/cv05_sum_trim/text - text - text allow_variable_data_keys: false max_cache_size: 0.0 max_cache_fd: 32 valid_max_cache_size: null optim: adam optim_conf: lr: 0.001 scheduler: reducelronplateau scheduler_conf: mode: min factor: 0.5 patience: 1 token_list: - - - '[hes]' - S - ▁THE - ▁TO - '''' - ▁AND - ▁YOU - ▁A - ▁IT - T - ▁THAT - ▁OF - ▁I - ▁IS - RE - ▁IN - ING - ▁WE - M - ▁GOING - ▁SO - ▁THIS - ▁YOUR - ▁ON - E - D - ▁BE - ▁CAN - N - Y - O - ER - ▁HAVE - ▁JUST - ▁FOR - ▁WITH - ▁DO - ED - ▁ARE - ▁WANT - ▁UP - R - LL - P - ▁ - L - B - ▁IF - C - ▁ONE - ▁S - ▁OR - A - ▁GO - ▁LIKE - ▁NOW - ▁HERE - VE - LE - U - ▁GET - ▁WHAT - ▁OUT - IN - W - ▁C - ▁LITTLE - ▁THERE - LY - ▁AS - ▁MAKE - I - ▁THEY - ▁MY - K - ▁THEN - ▁BUT - AL - G - ▁ALL - OR - ▁BACK - ▁NOT - ▁ABOUT - ▁RIGHT - ▁OUR - EN - ▁SOME - ▁DOWN - F - ▁WHEN - CH - ▁F - ▁HOW - AR - ▁WILL - ▁RE - CK - ▁G - ES - CE - ▁TAKE - ▁AT - ▁FROM - ▁WAY - TER - ▁SEE - RA - ▁USE - ▁REALLY - RI - TH - ▁TWO - ▁ME - ▁VERY - ▁E - ▁B - AT - ▁THEM - ▁DON - ▁AN - ▁BECAUSE - ▁MORE - RO - H - 'ON' - LI - ▁PUT - ▁ST - IL - ▁BIT - ▁START - ▁NEED - ▁INTO - UR - ▁TIME - ▁OVER - ▁W - ▁DE - ▁LOOK - ▁THESE - ▁LET - ▁GOOD - ▁ALSO - AN - ▁OFF - ▁HE - ▁KIND - ▁SIDE - ▁CO - ▁SURE - ▁AGAIN - ▁MA - ▁KNOW - IT - ▁WOULD - IC - ▁OTHER - LA - ▁P - ▁WHICH - '-' - IR - ▁LA - ▁HAND - EL - ▁LOT - ▁WHERE - ▁THREE - ▁PA - ION - LO - ▁KEEP - ▁SHOW - ▁THING - ▁FIRST - TE - ENT - ATE - ▁COME - AD - ▁GOT - NG - ▁NICE - ▁T - ET - ▁MO - ▁ANY - ▁ACTUALLY - ▁DIFFERENT - ▁SE - GE - ▁WORK - ▁THROUGH - ▁O - KE - V - ▁AROUND - ▁BA - PE - ▁HI - ▁BY - SH - ATION - ▁SU - ▁CA - ▁D - ▁LO - ▁HAS - ▁LI - ▁PLAY - Z - ▁ADD - ▁RO - ▁TA - AS - ▁FOUR - ▁CON - ▁THOSE - MP - NE - ▁SP - UT - ▁GIVE - ▁WELL - ▁BALL - TING - RY - X - ▁HO - INE - IVE - ▁NEXT - ▁PO - ▁STEP - ▁EVEN - TION - ▁MI - MENT - ▁CUT - ▁BO - ▁LINE - ▁MUCH - ▁THINGS - ▁TALK - UN - ▁PART - ▁WAS - ▁FA - ▁SOMETHING - PP - ANCE - ND - DI - ▁RA - AGE - ▁SAME - ▁EXPERT - ▁DOING - ▁LEFT - IST - ▁DI - ▁NO - RU - ME - TA - UL - TI - ▁VILLAGE - DE - ERS - ▁PEOPLE - ▁TURN - VER - ▁FL - ▁LEG - ▁ONCE - ▁COLOR - ▁PULL - ▁USING - VI - ▁WATER - ▁SHE - ▁TOP - ▁OKAY - ▁ANOTHER - ▁THEIR - ▁SAY - URE - ▁HA - ▁IMPORTANT - ▁PIECE - ▁FOOT - ▁TRA - ▁SC - ▁BODY - ▁SET - ▁POINT - ▁HELP - ▁TODAY - ▁BRING - ▁V - ▁END - MA - ▁CH - ▁MOST - ▁K - ▁AHEAD - ▁HER - OL - ▁SA - AM - IES - ▁THINK - ▁NAME - ▁TRY - ▁MOVE - ONE - ▁LE - ▁TOO - TO - UM - ▁PLACE - ▁COULD - ▁FIND - ▁FIVE - ▁ALWAYS - ID - TY - NT - ▁FEEL - ▁HEAD - ▁THAN - NA - ▁EX - ▁EYE - ITY - CI - OP - ▁SHOULD - ▁MIGHT - ▁HOLD - ▁CAR - AND - ▁GREAT - ▁RI - ▁BU - ▁HIGH - ▁OPEN - ▁BEFORE - US - ▁FRONT - ▁LONG - ▁TOGETHER - NI - ▁HAIR - ▁LIGHT - ▁TEN - ▁HIT - EST - OUS - ▁PRETTY - ▁TYPE - IP - CO - ▁FINGER - ▁JO - ▁UN - ▁PRO - ▁STRAIGHT - ▁BEHALF - ▁TI - ▁SIX - ▁CLEAN - ▁DIS - ▁DA - ▁POSITION - IGHT - ACT - ▁CHA - ▁PE - GG - AP - ▁MEAN - ▁COMP - FI - ▁KNEE - ▁CALLED - ▁HANDS - ▁PRE - ▁FORWARD - ▁AREA - ANT - ▁TE - ▁WA - ▁AFTER - ▁SMALL - ▁THROW - ▁EVERY - ▁SHOULDER - NC - PER - ▁MAYBE - ▁ABLE - ▁BASICALLY - ▁AM - ▁READY - ▁BOTTOM - IE - ▁HALF - FF - ▁BIG - ▁EACH - ▁PUSH - ▁EIGHT - ▁NEW - ▁DONE - ▁MAY - ▁GETTING - HO - ▁HIS - ▁HARD - ▁CLOSE - ALLY - ▁SECOND - ▁FEET - ICAL - ▁JA - ▁PAINT - ▁LEARN - ▁SOUND - HE - ▁ROLL - ▁ONLY - ▁DOESN - WA - ▁DRAW - ▁VI - ▁DID - ▁SHA - ▁CENTER - CU - ▁CLIP - ▁PI - ▁CARD - ▁INSIDE - ▁PERSON - ▁STILL - ▁MAKING - 'NO' - ▁EVERYTHING - . - ▁FUN - ARD - ▁REMEMBER - ▁AWAY - ATED - COM - ▁SEVEN - ▁BEEN - ▁MANY - ABLE - ▁DAY - ▁SIT - IZE - ▁REAL - ▁HIP - ▁BASIC - ▁KICK - ▁TU - ATING - ▁STICK - ▁FLAT - ▁WHO - END - HA - ▁EXP - ▁PICK - ▁MIX - ▁TRI - ▁BI - ▁WHOLE - ▁STRETCH - ▁BOTH - ▁PROBABLY - CA - ▁HIM - ▁STRING - ▁EDGE - ▁BASE - ▁COMING - UGH - ▁LIFT - ▁STA - ▁WORKING - ▁MU - ▁QUICK - ▁SOMETIMES - ▁HAPPEN - ▁YOURSELF - ▁TALKING - ▁DR - ▁TELL - ▁ANYTHING - ▁BRA - ▁LOOKING - ▁SLOW - ▁NE - ▁STAND - NER - ▁COMES - ▁GOES - ISE - BE - ▁USED - ▁UNDER - ▁BETWEEN - ▁HU - ▁CREATE - ▁NA - ▁USUALLY - ▁ARM - ▁DRY - ▁RUN - LING - ▁BRUSH - ▁COVER - ▁HEAR - ▁DOES - ▁STAY - ▁EN - ▁FOLD - ▁CHANGE - ▁LAST - ▁EASY - ▁US - ▁PER - ▁FACE - ▁EAR - ▁TIGHT - ▁FE - ▁PIN - ▁MAN - ▁BETTER - ▁CALL - ▁PRI - ▁BEST - ▁KI - ▁COUPLE - ▁WHILE - ▁SHAPE - ▁GAME - IV - ▁SHOT - ▁PAPER - ▁OWN - ▁ALRIGHT - ▁HAD - TIC - ▁BREATH - ▁TOOL - '2' - ▁ENOUGH - ▁COURSE - ▁SKIN - ▁SPIN - ▁VA - ▁ARMS - ▁TEA - ▁BREAK - ▁DOG - ▁1 - QUE - ▁DROP - ▁NUMBER - IG - ▁RED - ▁NOTE - ▁WEIGHT - WARD - ▁PLAYING - ▁FINISH - ▁MINUTE - ▁R - ▁PRESS - ▁EITHER - ▁CHE - ▁PU - BER - ▁FEW - ▁SIZE - ▁MADE - ▁LEAVE - ▁GA - ▁ALREADY - ▁GUY - ▁FAR - ▁HOME - ▁BAR - UP - ▁GRAB - ▁MARK - ▁WHITE - ▁PROPER - ▁CAUSE - ▁OK - ▁ART - HI - ▁SORT - ▁EXERCISE - ▁LOWER - PORT - ▁PLANT - ▁BOARD - ▁CASE - ▁YEAR - CENT - ▁DU - ▁CHECK - ▁WHATEVER - ▁OIL - ▁IDEA - ▁SIMPLE - ▁PRACTICE - ▁FAST - '0' - ▁CONTROL - ▁J - ▁KEY - ▁MIDDLE - ▁FULL - ▁GLASS - ▁OUTSIDE - ▁LOW - ▁REST - ▁STUFF - ▁ACT - ▁UNTIL - ▁BLACK - ▁POP - ▁CLICK - ▁HOLE - ▁Z - ▁COUNT - ▁POT - ▁ALLOW - ▁HAVING - ▁TRYING - ▁MUSCLE - ▁GU - ▁BOX - ▁NOTICE - ▁EXAMPLE - UND - ▁ALONG - FUL - ISH - ▁STORE - ▁LU - ▁FLOOR - ▁MOVING - ▁LARGE - ▁STOP - ▁PH - ▁WALK - '5' - ▁QU - ▁TECHNIQUE - ▁SOFT - ▁GROUND - ▁JUMP - ▁JU - ▁FILL - ▁WHY - ▁BUY - ▁GREEN - ▁WALL - ▁HEEL - NESS - ▁LEVEL - ▁UNDERNEATH - ▁PATTERN - ▁BEHIND - ▁OLD - ▁TIP - ▁COMPLETE - ▁WON - ▁TEACH - ▁FIT - ▁NECK - ▁REMOVE - ▁TRICK - ▁MOVEMENT - ▁TOWARDS - ▁PARTICULAR - ▁CHI - ▁EFFECT - J - ▁FREE - ▁ACROSS - ▁BEND - ▁SAFE - ▁SLIDE - ▁PROBLEM - ▁BLOCK - ▁PAN - ▁NATURAL - ▁TOUCH - ▁CHILD - LINE - ▁CROSS - ▁REASON - '4' - ▁POWER - ▁APPLY - ▁FOLLOW - ▁DESIGN - ▁SPACE - ▁ORDER - ▁WOOD - ▁RID - '3' - ▁COOK - ▁BEGIN - ▁WATCH - ▁STYLE - QUA - ▁PRODUCT - ▁TAKING - ▁PUTTING - ▁EXHALE - ▁THOUGH - ▁DEEP - IAN - ▁REACH - ▁FOOD - ▁ALMOST - ▁COOL - ▁SECTION - ▁SAID - ▁ANGLE - ▁MUSIC - ▁RELAX - ▁CORNER - ▁DARK - ▁CHORD - ▁ESPECIALLY - ▁SCALE - ▁WARM - ▁WITHOUT - ▁WHEEL - ▁SEGMENT - ▁TABLE - ▁BOOK - ▁PASS - ▁ELBOW - ▁ROUND - ▁INHALE - ▁SMOOTH - ▁ROOM - / - ▁NINE - ▁SHORT - ▁MEASURE - ▁LESS - ▁TWIST - ▁BALANCE - ▁PROCESS - ▁SWITCH - ▁GENERAL - ▁CLAY - ▁CERTAIN - ▁NEVER - ▁BLUE - ▁CUP - ▁HOUSE - ▁EXTRA - ▁MOTION - ▁PRESSURE - ▁FIRE - ▁SIMPLY - ▁DOUBLE - ▁TWENTY - ▁CATCH - ▁BECOME - ▁BUILD - ▁SPEED - ▁TRANS - ▁DRUM - ▁CHEST - ▁PICTURE - ▁LENGTH - ▁CONTINUE - ▁COMFORTABLE - ▁FISH - ▁PHOTO - ▁LOOSE - ▁SKI - ▁LIFE - ▁DEGREE - ▁OPTION - ▁WORD - ▁SHARP - ▁SHOOT - ▁FOUND - ▁STRONG - ▁QUITE - ▁THIRD - ▁GLUE - ▁MIND - ▁DEFINITELY - ▁EASIER - GRAPH - ▁HOOK - ▁CLEAR - ▁POSE - ▁BUTTON - ▁CHOOSE - ▁THICK - ▁SYSTEM - ▁PERFECT - ▁BEAUTIFUL - ▁SPOT - ▁GROW - ▁SIGN - ▁ELSE - ▁CONNECT - ▁SELECT - ▁PUNCH - ▁DIRECTION - ▁WRAP - ▁RELEASE - QUI - SIDE - ▁CAREFUL - ▁VIDEO - ▁INSTEAD - ▁CIRCLE - ▁WIRE - ▁NOSE - ▁AMOUNT - ▁FOCUS - ▁NORMAL - ▁MAJOR - ▁WHETHER - ▁SURFACE - ▁THUMB - ▁DRIVE - ▁SCREW - ▁POSSIBLE - ▁OBVIOUSLY - ▁COMMON - ▁REGULAR - ▁ADJUST - ▁WIDE - ▁BLADE - ▁FRET - ▁RECOMMEND - ▁BOWL - BOARD - ▁IMAGE - ▁DEPENDING - ▁PROTECT - ▁CLOTH - ▁HEALTH - ▁WRIST - ▁CLUB - ▁DRINK - ▁SINCE - ▁FRIEND - '00' - ▁RUNNING - ▁ITSELF - ▁RECORD - ▁SWING - ▁DIRECT - ▁MATERIAL - ▁YO - ▁LEAST - ▁EXACTLY - ▁BEGINNING - ▁SLIGHTLY - ▁TREAT - ▁CAMERA - ▁QUARTER - ▁WINDOW - '8' - ▁SOMEBODY - ▁BURN - ▁DEMONSTRATE - ▁DIFFERENCE - ▁COMPUTER - IBLE - ▁SHOE - ▁PERFORM - ▁SQUARE - ▁CONSIDER - ▁DRILL - ▁TEXT - ▁FILE - ▁RUB - ▁FABRIC - ▁HUNDRED - ▁GRIP - ▁CHARACTER - ▁SPECIFIC - ▁KNOT - ▁CURL - ▁STITCH - ▁BLEND - ▁FRAME - ▁THIRTY - '1' - ▁HORSE - ▁ATTACH - ▁GROUP - ▁STROKE - ▁GUITAR - ▁APART - ▁MACHINE - ▁CLASS - ▁COMB - ▁ROOT - ▁HELLO - ▁ENERGY - ▁ATTACK - ▁CORRECT - ▁EXTEND - ▁MINOR - ▁PROFESSIONAL - ▁MONEY - ▁STRIP - ▁FLAVOR - ▁EVERYBODY - ▁RULE - ▁DIFFICULT - ▁PROJECT - ▁DISCUSS - ▁FIGURE - ▁HOWEVER - ▁FINAL - ▁STRENGTH - ▁ENTIRE - ▁FIELD - ▁CONTACT - ▁SUPPORT - ▁PALM - ▁SERIES - ▁ENJOY - '6' - ▁WORLD - ▁DECIDE - ▁SPEAK - ▁SEVERAL - ▁WRITE - ▁PROGRAM - ABILITY - ▁KNIFE - ▁PLASTIC - ▁ORGAN - '7' - ▁UNDERSTAND - ▁FIFTEEN - ▁FLEX - ▁INFORMATION - ▁TWELVE - ▁DETAIL - ▁STRIKE - ▁ACTUAL - ▁SPRAY - ▁LOCAL - ▁MOUTH - ▁NIGHT - ▁VEHICLE - ▁OPPOSITE - ▁SCHOOL - '9' - ▁QUESTION - ▁SPECIAL - ▁BIGGER - ▁DEVELOP - ▁PEPPER - ▁PREFER - Q - '%' - ']' - '[' - '&' - ',' - _ - '#' - '=' - '@' - + - '*' - $ - '~' - init: null input_size: null ctc_conf: ignore_nan_grad: true model_conf: ctc_weight: 0.0 lsm_weight: 0.15 length_normalized_loss: false use_preprocessor: true token_type: bpe bpemodel: data/en_token_list/bpe_unigram1000/bpe.model non_linguistic_symbols: data/nlsyms cleaner: null g2p: null speech_volume_normalize: null rir_scp: null rir_apply_prob: 1.0 noise_scp: null noise_apply_prob: 1.0 noise_db_range: '13_15' frontend: default frontend_conf: n_fft: 512 hop_length: 256 fs: 16k specaug: specaug specaug_conf: apply_time_warp: true time_warp_window: 5 time_warp_mode: bicubic apply_freq_mask: true freq_mask_width_range: - 0 - 30 num_freq_mask: 2 apply_time_mask: true time_mask_width_range: - 0 - 40 num_time_mask: 2 normalize: global_mvn normalize_conf: stats_file: exp/asr_stats_raw_vid_sum/train/feats_stats.npz preencoder: null preencoder_conf: {} encoder: conformer encoder_conf: output_size: 512 attention_heads: 8 linear_units: 2048 num_blocks: 12 dropout_rate: 0.1 positional_dropout_rate: 0.1 attention_dropout_rate: 0.1 input_layer: conv2d normalize_before: true macaron_style: true pos_enc_layer_type: abs_pos selfattention_layer_type: lf_selfattn activation_type: swish use_cnn_module: true cnn_module_kernel: 31 attention_windows: - 40 - 40 - 40 - 40 - 40 - 40 - 40 - 40 - 40 - 40 - 40 - 40 attention_dilation: - 1 - 1 - 1 - 1 - 1 - 1 - 1 - 1 - 1 - 1 - 1 - 1 attention_mode: tvm decoder: transformer decoder_conf: attention_heads: 4 linear_units: 512 num_blocks: 6 dropout_rate: 0.15 positional_dropout_rate: 0.15 self_attention_dropout_rate: 0.15 src_attention_dropout_rate: 0.15 required: - output_dir - token_list version: 0.10.0 distributed: true ```
Please cite the following paper if you use this recipe: ```BibTex @misc{sharma2022speech, title={Speech Summarization using Restricted Self-Attention}, author={Roshan Sharma and Shruti Palaskar and Alan W Black and Florian Metze}, year={2022}, eprint={2110.06263}, archivePrefix={arXiv}, primaryClass={cs.CL} } ``` ### Citing ESPnet ```BibTex @inproceedings{watanabe2018espnet, author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai}, title={{ESPnet}: End-to-End Speech Processing Toolkit}, year={2018}, booktitle={Proceedings of Interspeech}, pages={2207--2211}, doi={10.21437/Interspeech.2018-1456}, url={http://dx.doi.org/10.21437/Interspeech.2018-1456} } ``` or arXiv: ```bibtex @misc{watanabe2018espnet, title##3={ESPnet: End-to-End Speech Processing Toolkit}, author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai}, year={2018}, eprint={1804.00015}, archivePrefix={arXiv}, primaryClass{cs.CL} ```