[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84048-en":3,"doc-seo-84048-105":29,"detail-sidebar-cat-0-en-105":82},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},84048,1374391975076,"Riley","https://ap-avatar.wpscdn.com/avatar/14000253ca4ec9f6853?x-image-process=image/resize,m_fixed,w_180,h_180&k=1783305029341752051",8,"Research & Report","Fréchet Distance Loss on Speech Representations for Text-to-Speech Synthesis","Few-step diffusion and flow-matching text-to-speech (TTS) models are often trained with local objectives that do not verify whether sampled speech follows the distribution of high-quality speech. The document introduces Speech Representation Fréchet Distance loss (SR-FD), a training-time distributional regularizer for tokenizer-free flow-matching autoregressive TTS. During fine-tuning, SR-FD matches the mean and covariance of frozen Whisper and CTC speech representations from generated audio to offline reference statistics. On Seed-TTS English, four-step SR-FD fine-tuning reduces WER from 2.23% to 1.41% while recovering quality and speaker similarity; listening tests show no reliable preference.","Fréchet Distance Loss on Speech Representations for  \nText-to-Speech Synthesis  \nHo-Lam Chung 1 , Kuan-Po Huang2 , Bo-Ru Lu†, Hung-yi Lee 1  \n1 Graduate Institute of Communication Engineering, National Taiwan University, Taipei, Taiwan  \n2 Graduate Institute of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan  \n†Amazon  \narXiv :2607 .06027v2 [ cs . SD] 11 Jul 2026  \nAbstract—Few-step diffusion and flow-matching text-to-speech (TTS) models are usually trained with local objectives, such as conditional flow matching, reconstruction, and stop prediction. These losses provide stable optimization, but they never ask whether sampled speech follows the distribution of high-quality speech. We propose Speech Representation Fréchet Distance loss (SR-FD), a training-time distributional regularizer for tokenizerfree flow-matching autoregressive TTS. During fine-tuning, the model synthesizes speech with the same few-step sampler used at deployment, and SR-FD matches the mean and covariance of frozen Whisper and CTC features of this speech to reference statistics computed offline from three complementary content targets. The loss requires no discriminator and no inferencetime computation. On Seed-TTS English, four-step SR-FD finetuning reduces word error rate (WER) from the original four-step VoxCPM2 baseline’s 2.23% to 1.41%, a 36.5% relative reduction, and also surpasses the original ten-step baseline at 1.74%; both gains are significant under an utterance-level paired bootstrap. Speaker similarity and objective quality proxies are close to the ten-step level, and a blinded listening test finds no reliable listener preference against the ten-step baseline. An error analysis attributes the gain to fewer content substitutions across all prompt lengths. We frame SR-FD as an intelligibility regularizer for fewstep TTS, not a general quality objective.  \nIndex Terms—text-to-speech, flow matching, Fréchet distance, speech generation, LoRA, VoxCPM2  \nI. INTRODUCTION  \nDiffusion and flow-matching TTS systems achieve high synthesis quality [1]–[4], but they require multiple numerical integration steps at inference, imposing deployment latency. Reducing the number of flow-matching integration steps is therefore a practical strategy for real-time use. Standard training objectives, however, operate locally: the flow-matching loss supervises a per-frame velocity field, reconstruction losses target individual acoustic frames, and the stop-prediction loss governs utterance length termination [5] . Teacher-distillation methods [6], [7] reduce the trajectory-level step-count gap, but they do not directly constrain the population distribution of complete utterances that the compressed sampler produces. A model may therefore achieve low training loss while the distribution of its sampled outputs drifts from that of full-step inference in content, duration, prosody, and acoustic quality. This issue is especially visible in tokenizer-free flow-matching autoregressive TTS models. VoxCPM2 [5] generates continuous  \nacoustic representations rather than discrete codec tokens,†This work is unrelated to the author’s position at Amazon.  \navoiding quantization artifacts of discrete-based systems, but when its sampler is compressed to four steps, the generated distribution can drift from the one generated by more sampling steps. In our Seed-TTS English evaluation [8], the original VoxCPM2 four-step sampling baseline has a substantially worse word error rate (WER) than the original ten-step baseline, showing that the low-step sampler is not merely a faster version of the same model. We propose Speech Representation Fréchet Distance loss (SR-FD), a training-time distributional regularizer for tokenizer-free flow-matching autoregressive TTS. During fine-tuning, SR-FD synthesizes speech with the same flowmatching sampler used at deployment, extracts frozen speech representations from the generated audio, and matches the mean and co","cbCaifgQaEOnsvML","https://ap.wps.com/l/cbCaifgQaEOnsvML","pdf",280453,1,7,"English","en",105,"# Introduction\n## Few-step flow-matching motivation\n## Limits of local training objectives\n## Proposed SR-FD and distribution matching","[{\"question\":\"What performance gains are reported for Seed-TTS English?\",\"answer\":\"Four-step SR-FD fine-tuning reduces word error rate (WER) from 2.23% to 1.41%, a 36.5% relative reduction, and it also surpasses the original ten-step baseline at 1.74%. Listening tests show no reliable listener preference against the ten-step baseline.\"}]",1784192229,18,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":77,"head_meta":79,"extra_data":81,"updated_unix":27},"frechet-distance-loss-on-speech-representations-for-text-to-speech-synthesis","",{"@graph":35,"@context":76},[36,53,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/frechet-distance-loss-on-speech-representations-for-text-to-speech-synthesis/84048/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70],{"name":71,"@type":72,"acceptedAnswer":73},"What performance gains are reported for Seed-TTS English?","Question",{"text":74,"@type":75},"Four-step SR-FD fine-tuning reduces word error rate (WER) from 2.23% to 1.41%, a 36.5% relative reduction, and it also surpasses the original ten-step baseline at 1.74%. Listening tests show no reliable listener preference against the ten-step baseline.","Answer","https://schema.org",{"og:url":51,"og:type":78,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":80,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":83},[84,88,92,96,101,106,110,113,118,121,125],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":102,"doc_module":4,"doc_module_name":45,"category_name":103,"show_sort_weight":104,"slug":105},6,"Technology",50,"technology",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":111,"slug":112},30,"research-report",{"id":114,"doc_module":4,"doc_module_name":45,"category_name":115,"show_sort_weight":116,"slug":117},9,"Religion & Spirituality",20,"religion-spirituality",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":119,"show_sort_weight":116,"slug":120},"World Cup","world-cup",{"id":122,"doc_module":4,"doc_module_name":45,"category_name":123,"show_sort_weight":122,"slug":124},10,"Lifestyle","lifestyle",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":127,"show_sort_weight":97,"slug":128},19,"General","general"]