{"id":"https://openalex.org/W4297841445","doi":"https://doi.org/10.21437/interspeech.2022-413","title":"Accelerating Inference and Language Model Fusion of Recurrent Neural Network Transducers via End-to-End 4-bit Quantization","display_name":"Accelerating Inference and Language Model Fusion of Recurrent Neural Network Transducers via End-to-End 4-bit Quantization","publication_year":2022,"publication_date":"2022-09-16","ids":{"openalex":"https://openalex.org/W4297841445","doi":"https://doi.org/10.21437/interspeech.2022-413"},"language":"en","primary_location":{"id":"doi:10.21437/interspeech.2022-413","is_oa":false,"landing_page_url":"https://doi.org/10.21437/interspeech.2022-413","pdf_url":null,"source":{"id":"https://openalex.org/S4363604309","display_name":"Interspeech 2022","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Interspeech 2022","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":false,"oa_status":"closed","oa_url":null,"any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5026800272","display_name":"Andrea Fasoli","orcid":"https://orcid.org/0000-0001-6892-5139"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Andrea Fasoli","raw_affiliation_strings":["IBM Research, USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research, USA","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101406569","display_name":"Chia\u2010Yu Chen","orcid":"https://orcid.org/0000-0001-5542-7149"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chia-Yu Chen","raw_affiliation_strings":["IBM Research, USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research, USA","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5042998510","display_name":"Maur\u00edcio Serrano","orcid":"https://orcid.org/0000-0003-0250-5881"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mauricio Serrano","raw_affiliation_strings":["IBM Research, USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research, USA","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5010094713","display_name":"Swagath Venkataramani","orcid":"https://orcid.org/0000-0002-0470-6364"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Swagath Venkataramani","raw_affiliation_strings":["IBM Research, USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research, USA","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5079994647","display_name":"George Saon","orcid":"https://orcid.org/0009-0004-6837-5009"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"George Saon","raw_affiliation_strings":["IBM Research, USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research, USA","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5102014291","display_name":"Xiaodong Cui","orcid":"https://orcid.org/0000-0003-4865-1307"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiaodong Cui","raw_affiliation_strings":["IBM Research, USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research, USA","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5003725957","display_name":"Brian Kingsbury","orcid":"https://orcid.org/0000-0002-1343-6837"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Brian Kingsbury","raw_affiliation_strings":["IBM Research, USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research, USA","institution_ids":[]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5101429456","display_name":"Kailash Gopalakrishnan","orcid":"https://orcid.org/0000-0002-8952-0875"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kailash Gopalakrishnan","raw_affiliation_strings":["IBM Research, USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"IBM Research, USA","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":5,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"2038","last_page":"2042"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10201","display_name":"Speech Recognition and Synthesis","score":0.9997000098228455,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10201","display_name":"Speech Recognition and Synthesis","score":0.9997000098228455,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.996999979019165,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.9934999942779541,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/end-to-end-principle","display_name":"End-to-end principle","score":0.7741150856018066},{"id":"https://openalex.org/keywords/quantization","display_name":"Quantization (signal processing)","score":0.6502683758735657},{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.6416000127792358},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5489516854286194},{"id":"https://openalex.org/keywords/artificial-neural-network","display_name":"Artificial neural network","score":0.5143886208534241},{"id":"https://openalex.org/keywords/speech-recognition","display_name":"Speech recognition","score":0.41372743248939514},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.3288046717643738},{"id":"https://openalex.org/keywords/algorithm","display_name":"Algorithm","score":0.3028959035873413}],"concepts":[{"id":"https://openalex.org/C74296488","wikidata":"https://www.wikidata.org/wiki/Q2527392","display_name":"End-to-end principle","level":2,"score":0.7741150856018066},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.6502683758735657},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6416000127792358},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5489516854286194},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.5143886208534241},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.41372743248939514},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3288046717643738},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3028959035873413}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.21437/interspeech.2022-413","is_oa":false,"landing_page_url":"https://doi.org/10.21437/interspeech.2022-413","pdf_url":null,"source":{"id":"https://openalex.org/S4363604309","display_name":"Interspeech 2022","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Interspeech 2022","raw_type":"proceedings-article"}],"best_oa_location":null,"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":27,"referenced_works":["https://openalex.org/W1828163288","https://openalex.org/W2407080277","https://openalex.org/W2803431233","https://openalex.org/W2936774411","https://openalex.org/W2937780860","https://openalex.org/W2962760690","https://openalex.org/W2975044525","https://openalex.org/W2981680926","https://openalex.org/W3005538384","https://openalex.org/W3008037978","https://openalex.org/W3015190365","https://openalex.org/W3015194534","https://openalex.org/W3016234571","https://openalex.org/W3017746288","https://openalex.org/W3036601975","https://openalex.org/W3037288590","https://openalex.org/W3048889251","https://openalex.org/W3094667432","https://openalex.org/W3097522836","https://openalex.org/W3163300396","https://openalex.org/W3163793923","https://openalex.org/W3172967059","https://openalex.org/W3196364802","https://openalex.org/W4224918069","https://openalex.org/W4226033575","https://openalex.org/W4289478342","https://openalex.org/W4302760292"],"related_works":["https://openalex.org/W3179968364","https://openalex.org/W1999612375","https://openalex.org/W2938107654","https://openalex.org/W2151749779","https://openalex.org/W3008587939","https://openalex.org/W3196421258","https://openalex.org/W2055243143","https://openalex.org/W4387301579","https://openalex.org/W2763956190","https://openalex.org/W4400488565"],"abstract_inverted_index":{"We":[0],"report":[1],"on":[2,83],"aggressive":[3],"quantization":[4,49,98],"strategies":[5,99],"that":[6,47,51,96],"greatly":[7],"accelerate":[8],"inference":[9],"of":[10,58,72,93,118,143,153,168],"Recurrent":[11],"Neural":[12],"Network":[13],"Transducers":[14],"(RNN-T).We":[15],"use":[16],"a":[17,113,130,148],"4":[18],"bit":[19],"integer":[20],"representation":[21],"for":[22,105,137],"both":[23],"weights":[24],"and":[25,27,40,43,112,161],"activations":[26],"apply":[28],"Quantization":[29],"Aware":[30],"Training":[31],"(QAT)":[32],"to":[33,54,63,121,135],"retrain":[34],"the":[35,55,59,69,90,122,138,155,169,176],"full":[36,114,123],"model":[37,115],"(acoustic":[38],"encoder":[39],"language":[41],"model)":[42],"achieve":[44,64],"near-iso-accuracy.We":[45],"show":[46,95],"customized":[48],"schemes":[50],"are":[52,61],"tailored":[53],"local":[56],"properties":[57],"network":[60],"essential":[62],"good":[65],"performance":[66],"while":[67,108],"limiting":[68],"computational":[70,91],"overhead":[71],"QAT.Density":[73],"ratio":[74,117],"Language":[75],"Model":[76],"fusion":[77],"has":[78],"shown":[79],"remarkable":[80],"accuracy":[81],"gains":[82,170],"RNN-T":[84,141],"workloads":[85],"but":[86],"it":[87],"severely":[88],"increases":[89],"cost":[92],"inference.We":[94],"our":[97],"enable":[100],"using":[101],"large":[102],"beam":[103],"widths":[104],"hypothesis":[106],"search":[107],"achieving":[109],"streaming-compatible":[110],"runtimes":[111],"compression":[116],"7.6\u00d7":[119],"compared":[120],"precision":[124],"model.Via":[125],"hardware":[126],"simulations,":[127],"we":[128,165],"estimate":[129],"3.4\u00d7":[131],"acceleration":[132],"from":[133],"FP16":[134],"INT4":[136],"end-to-end":[139],"quantized":[140],"inclusive":[142],"LM":[144,173],"fusion,":[145,174],"resulting":[146],"in":[147],"Real":[149],"Time":[150],"Factor":[151],"(RTF)":[152],"0.06.On":[154],"NIST":[156],"Hub5":[157,159],"2000,":[158],"2001,":[160],"RT-03":[162],"test":[163],"sets,":[164],"retain":[166],"most":[167],"associated":[171],"with":[172],"improving":[175],"average":[177],"WER":[178],"by":[179],">1.5%.":[180]},"counts_by_year":[{"year":2025,"cited_by_count":2},{"year":2024,"cited_by_count":1},{"year":2023,"cited_by_count":2}],"updated_date":"2026-07-14T23:27:15.235271","created_date":"2025-10-10T00:00:00"}