Top Speech And Audio Data Annotation Companies In 2027
Dominic Tarn
Head Writer - ReVerb
11.01.2026
Meta Description: Compare the top speech and audio data annotation companies in 2027 for ASR transcription, diarization, multilingual speech collection and voice agent data.
Speech models fail in predictable places. Word error rates climb on accented English, overlapping talkers, noisy call center lines, code-switched conversations and low-resource languages, and the gap usually traces back to training data that was too clean, too scripted or too narrow. Teams building ASR engines, voice agents and text-to-speech systems therefore spend a large share of their data budget on audio that looks like the real world: spontaneous dialogue, real devices, real background noise and speakers who were recruited for the demographics the product will actually serve.
Speech and audio data annotation companies handle that work. The strongest providers combine three things: a vetted contributor network that can record native speakers at scale, linguists who can transcribe to strict conventions with timestamps and speaker labels, and quality control that holds up across dozens of languages at once. Many also sell off-the-shelf speech corpora, which can shorten a project by months when a ready-made dataset fits. If you need a wider view of the category beyond audio, our guide to the top data annotation and labeling companies covers image, text and multimodal vendors as well.
The ten companies below were selected because each publishes a verifiable speech-specific offering: custom speech collection, ASR transcription, speaker diarization, sound event labeling, TTS voice data, emotion and intent tagging, or voice agent evaluation. Figures come from each company’s own site or press releases and were checked in October 2026.
Speech And Audio Data Annotation Companies At A Glance
Appen is one of the longest-running providers in the field, citing 30 years of AI data experience. The company is listed on the ASX under the ticker APX, with corporate headquarters in Sydney and a US headquarters in Kirkland, Washington. In a May 2026 announcement, Appen described a crowd of more than one million contributors, coverage of more than 500 global locales and speech data spanning more than 235 languages. It holds SOC 2 and ISO 27001 certification.
Appen’s speech line now covers the full voice stack. Its multimodal and speech AI offering lists multilingual ASR and conversational speech corpora, expressive text-to-speech annotation and voice agent evaluation, alongside custom data collection. For teams that want to start from existing data, Appen’s off-the-shelf catalog lists 646 datasets across nine categories, with speech sets covering read speech, conversational speech and contact center recordings. One published case study describes an in-car speech program that ran across more than 20 languages for over ten years.
Appen also contributes to public benchmarking. In May 2026 it supplied a suite of private English audio datasets to Hugging Face’s Open ASR Leaderboard, adding scripted, conversational and US versus non-US accent word error rate measures that are harder to overfit than public test sets. For ASR teams, that work signals where Appen’s data focus sits: conversational speech and accent coverage, the two areas where production models most often underperform. Appen fits large programs that need many languages delivered in parallel under one vendor.
2. LXT — 1,000+ Language Locales And ISO 27001:2022 Facilities
LXT was founded in 2010 by CEO Mohammad Omar and is headquartered in Canada, with offices in the United States, the UK, Egypt, India, Romania, Turkey, Germany and Australia. The company is part of the LXT Group, which also includes the crowdsourcing provider clickworker. LXT reports more than 10 million vetted global contributors, more than 250,000 domain experts and over 1,000 in-house specialists, with operations in more than 150 countries.
Language breadth is LXT’s clearest differentiator. The company states coverage of more than 1,000 language locales and dialects, including low-resource languages and niche dialects that many vendors cannot staff. Its speech services include audio data collection, speech-to-text transcription for short utterances and long-form audio, speaker diarization, timestamps and structured metadata, audio annotation and speech data evaluation. Transcription conventions can be tailored to a domain or project, which matters for teams that need consistent tagging of disfluencies, noise or named entities.
LXT also offers post-editing of ASR output, where human experts review automated transcripts and correct errors. That is a practical option for teams that already have a capable model and need to scale labeled data at lower cost than full manual transcription. For call center use cases, LXT converts recorded conversations into structured text for analytics and compliance work. The company holds ISO 27001:2022 certification and operates five ISO 27001-certified secure facilities, and it reports a 99%+ data acceptance rate on enterprise projects.
3. Defined.ai — 2 Million+ Hours Of AI-Ready Speech Data
Defined.ai was founded in Seattle in 2015 by CEO Daniela Braga and launched its first speech and NLP data platform in 2016 under the name DefinedCrowd. It keeps headquarters in Seattle and an R&D center in Lisbon, Portugal. The company reports 1.6 million+ experts worldwide, more than 120 customers and work across more than 150 markets.
The company’s main draw for speech teams is its marketplace. Defined.ai states it holds more than two million hours of AI-ready speech data across more than 500 languages and dialects, in more than eight data types including scripted speech, podcasts and call center conversations. Buyers can browse and license datasets for ASR and TTS training rather than commissioning collection from scratch, which can cut weeks from a project when the language and domain already exist in the catalog.
Defined.ai also runs custom work. Its custom data collection service produces TTS training data with voice talent reading phonetically balanced scripts, and its evaluation service has human raters assess TTS pronunciation, naturalness and context, available through an API or as a managed service. Supporting tools include phonetic balancing, grapheme-to-phoneme conversion and signal-to-noise checks. On governance, Defined.ai holds ISO 42001 certification for AI management systems, plus ISO 27001 and ISO 27701, and states GDPR compliance. That certification set suits buyers whose procurement teams ask for documented responsible AI controls on licensed data.
4. DataOcean AI — 1,600+ Off-The-Shelf Training Datasets
DataOcean AI has been serving customers since 2005, when it started out providing TTS and ASR data services. Its US headquarters is in Spokane, Washington, and it reports collection resources in more than 70 countries. The company announced a brand upgrade at ICASSP 2024 and says it has supplied data products and services to more than 1,000 internet companies, AI enterprises and academic institutes.
The company is best known for scale in prepackaged speech corpora. According to its September 2024 Interspeech announcement, DataOcean AI offers more than 1,600 off-the-shelf training datasets, over 1,100 of them speech datasets. Its Massively Multilingual Speech Corpus, built for speech foundation models, covers more than 100 languages and totals 259,672 hours recorded by 215,891 speakers. Individual sets can be large as well: its desktop Mandarin speech recognition corpus alone runs to about 5,868 hours.
For custom programs, DataOcean AI states that its speech collection and transcription cover more than 190 languages and dialects, supported by a recording network in over 200 languages. TTS buyers can draw on more than 10,000 voice actors and over 1,000 voiceover styles. Annotation runs on the company’s DOTS data processing platform, which it says includes more than 200 algorithms. DataOcean AI holds ISO 9001, ISO 27001 and ISO 27701 certification. It suits foundation model teams that need broad Asian and multilingual coverage quickly, and buyers who prefer to evaluate ready-made data before committing to collection.
5. Welo Data — 155+ Locales With Diarized, Timestamp-Aligned Transcription
Welo Data is the AI training data division of Welocalize, the language services company headquartered at 136 Madison Avenue in New York. Welocalize launched Welo Data as a dedicated brand in July 2024, drawing on more than 25 years of language work. Welo Data reports more than 500,000 verified expert evaluators and contributors, with native-speaker networks in over 100 countries, and verifies contributors through its NIMO workforce integrity platform.
Welo Data’s voice AI offering is one of the most complete service menus on this list. It covers ASR training data in scripted, spontaneous and conversational formats; TTS data production; phonetic, prosodic and semantic annotation; verbatim and normalized transcription with speaker diarization and timestamp alignment; emotion and intent tagging; and human perceptual evaluation of naturalness, fluency and intelligibility. That range lets one vendor handle both the training data and the listening tests used to judge a model’s output.
Coverage spans more than 155 locales, which Welo Data says includes low-resource and endangered varieties. The company is ISO 27001 certified and operates 14 secure delivery facilities worldwide, which is relevant for projects involving sensitive recordings that cannot leave a controlled environment. Welo Data states that typical pilot datasets are delivered in under two weeks, a useful benchmark when comparing vendors on how quickly a trial can begin. It fits teams building both recognition and synthesis that want transcription, labeling and evaluation from a single linguistic workforce.
6. Shaip — Off-The-Shelf Accented English And Call Center Audio
Shaip began in healthcare and medical transcription before shifting to AI training data after a 2019 project with a Fortune 10 company. It has operated as an AI data business since 2019 and in February 2026 became part of Ubiquity Global Services, which is headquartered in New York; Shaip continues to operate as a specialized AI data platform within Ubiquity. The company reports 12 delivery centers across four continents and a network of more than 30,000 contributors.
Shaip’s audio annotation menu is detailed. It covers transcription with word-level timestamps in verbatim, non-verbatim and phonetic forms; speaker diarization with optional gender, age-band and language tags; acoustic event and environmental sound classification; multi-label annotation for overlapping sources; intent, sentiment and dialect layers for utterances; emotion and voice-style tagging; and music classification. For generative voice work it supports RLHF preference ranking on audio, TTS dataset preparation and voice cloning corpora. Annotation covers more than 150 languages and dialects.
The off-the-shelf catalog is the reason many ASR teams look at Shaip. It lists call center audio by accent and variety, including African American Vernacular English (365 hours), Boston English (302 hours), Deep South English (473 hours), Hinglish (424 hours) and Singapore English (465 hours), along with large scripted monologue sets such as Hindi (2,867 hours). Collection formats include dual-channel dialogue, multi-party conversations, wake-word utterances and acoustic recordings in homes, offices and restaurants. Shaip runs its workflows under an ISO 27001-certified information security management system.
7. Magic Data — 200,000+ Hours Of Conversational And Read Speech
Magic Data Technology was founded in 2016 and is headquartered in Beijing. The company focuses on speech and language data for ASR, TTS and NLP, and holds ISO/IEC 27001, ISO/IEC 27701:2019 and CMMI Level 3 certification. Its data processing platform uses a human-in-the-loop setup with AI-assisted pre-labeling, automated quality inspection and pronunciation prediction for new words, which helps annotators maintain pronunciation dictionaries.
Conversational speech is Magic Data’s specialism. According to Speech Technology magazine, the company offers more than 200,000 hours of training data, about 140,000 hours of it conversational and 60,000 hours read speech, spanning Asian languages, English dialects and European languages. Magic Data’s R&D center reported that, in an internal test, models trained on 3,000 hours of conversational speech achieved up to 84% better word accuracy than models trained on the same amount of read speech across customer service, broadcasting and navigation scenarios. Buyers should treat that as a vendor result, but it reflects a widely held view that spontaneous dialogue is worth more per hour than scripted reading.
Today the company lists more than 250 ready-to-use datasets in over 60 languages, including duplex conversational data for full-duplex voice agents and voice asset datasets built for TTS and voice cloning. Datasets can be browsed through its Dataset Explorer, and the company also runs custom collection and annotation projects. Magic Data is a strong fit for teams that need Mandarin and other Asian language conversational data, or real two-way dialogue for voice agent training.
8. TELUS Digital — 4 Million+ Utterances Delivered For A 35-Language Voice Assistant
TELUS Digital is the digital services arm of Canadian telecom group TELUS, which completed the privatization of TELUS Digital on October 31, 2025, taking full ownership and delisting its shares from the TSX and NYSE. The company’s AI data solutions business reports more than 20 years in data work and an AI community of over one million contributors across more than 450 locations, with annotation available in more than 500 languages and dialects.
TELUS Digital’s speech credentials are clearest in its published voice assistant case study. For a multinational technology company, it collected more than four million audio prompts and utterances over five years, helping add 35 languages to the client’s voice assistant, with a focus on colloquial speech. More than 5,000 native-speaker contributors across 50 regions recorded on their own devices, verified through IP address tracking. TELUS Digital annotated the audio against guidelines developed with the client, led the creation of quality guidelines and performed QA on work delivered by other vendors on the same program.
Tooling includes Ground Truth Studio for labeling, Fine-Tune Studio and Experts Engine for custom fine-tuning data, and a dedicated data collection app, plus off-the-shelf audio and speech recognition benchmarking datasets. Its platforms are described as GDPR, SOC 2 and TISAX compliant, and its labeling facilities are ISO 27001 certified. TELUS Digital fits large enterprise voice programs that run over several years and need a vendor able to manage contributor verification and cross-vendor quality.
9. Andovar — 200+ Languages And 8 Recording Studios
Andovar is a localization and multilingual data company with offices in Singapore, Bangkok, Kolkata and Budapest, and it describes nearly two decades of work in multilingual services. Its data division supplies text, voice, speech, image and video data, and the company has held G2 Leader badges across several seasons from Spring 2025 to Winter 2026.
The voice data offering is built around two collection modes. Remote collection captures scripted and spontaneous speech on contributors’ own mobile devices or laptops, across varied environments and accents. Studio recording uses eight studios with high-end microphones and controlled acoustics, which suits neural TTS voices and speaker identification projects where recording quality matters more than ambient realism. Andovar reports more than 30,000 global voice contributors and supports scripted speech, conversational speech, spontaneous dialogue and voice-activated commands.
Language coverage is where Andovar stands out for smaller buyers. The company states it works in more than 200 languages and dialects, including over 40 low-resource and underserved languages, with reach into 120 countries. Its marketplace sells ready-made datasets alongside custom work, and Andovar reports more than 100,000 hours of professionally recorded, AI-ready voice data. Applications it lists include ASR, TTS and voice biometrics. Andovar is a practical choice for teams that need studio-grade voice recordings or a long-tail language that larger vendors staff only at high minimum volumes.
10. Sigma AI — Audio Annotation For Tone, Emotion And Intent
Sigma AI has more than 18 years of experience building AI data and evaluation programs and operates offices in Madrid, London and Miami. More than 2,500 people from over 50 nationalities work across the company, and it reports more than 70,000 vetted human experts on call, coverage of more than 700 languages and dialects, and over 30 enterprise clients in production. Sigma holds ISO 27001 and SOC 2 Type II certification and states GDPR compliance.
Sigma’s speech work is strongest on the annotation side, particularly the paralinguistic layers that standard transcription misses. In one published case study, a global technology company needed audio annotated for tone, emotion and intent, and Sigma assembled a specialized annotation team within two weeks. That kind of labeling feeds emotion recognition, expressive TTS and voice agents that need to detect frustration or urgency in a caller’s voice.
The company also handles multilingual transcription at speed. Another case study describes 2,000 hours of video transcribed by humans in 24 languages, with all 24 language teams launched at the same time. A third describes more than 1,000 dialect conversations collected in two months using automation and skilled linguists. Sigma fits teams that need careful human judgment on subjective audio labels, or many languages started in parallel, and that value a vendor with a large in-house staff rather than a purely crowd-based model.
How To Choose A Speech And Audio Data Annotation Company
Do You Need Custom Collection Or Off-The-Shelf Data?
Start by checking whether a licensed dataset already covers your language, domain and recording conditions. Defined.ai, DataOcean AI, Magic Data, Shaip and Appen all sell prepackaged corpora, and licensing can be far faster than recruiting speakers. Custom collection is the better option when you need a specific device, acoustic environment, wake word, demographic balance or domain vocabulary that catalogs do not hold. Many teams combine the two: off-the-shelf data for base training, then a smaller custom set that matches production conditions. For a wider look at how data costs fit into a model budget, see our guide to AI development cost.
How Does The Vendor Handle Accents, Dialects And Low-Resource Languages?
Ask for the number of native-speaker contributors the vendor can field in each target locale, not the headline language count. A provider may list hundreds of languages but staff some of them thinly. Ask how speakers are verified, since location and nativeness checks such as the IP verification TELUS Digital describes reduce the risk of non-native recordings slipping into a dataset. For accented English, request accent-labeled samples and confirm that accent metadata will be delivered with every file.
What Transcription Conventions And Labels Are Included?
Speech labeling varies widely. Confirm whether the vendor delivers verbatim or normalized transcripts, word-level or segment-level timestamps, speaker diarization, overlap marking, noise and non-speech event tags, and emotion or intent labels. Write your conventions into a style guide before the pilot, and ask the vendor to show inter-annotator agreement or word error rate on a gold set. If you also train language models on transcripts or run preference ranking on voice output, the vendors in our list of RLHF and LLM training data companies may overlap with your speech vendor.
Can The Vendor Evaluate Model Output, Not Just Label Data?
Voice products increasingly need human evaluation as much as training data. TTS teams need listening tests for naturalness and pronunciation, and voice agent teams need conversation-level scoring. Welo Data, Defined.ai and Appen all publish evaluation services. Choosing a vendor that can both produce data and evaluate results keeps rater pools and guidelines consistent across the training cycle.
How Is Sensitive Audio Secured?
Voice recordings are biometric data in many jurisdictions, and call center audio often contains personal information. Check for ISO 27001 certification, SOC 2 reports and GDPR handling, and ask whether sensitive work runs in secure facilities rather than on contributors’ home machines. LXT, Welo Data and TELUS Digital all describe ISO 27001-certified facilities. Confirm consent terms for every recorded speaker, especially for TTS and voice cloning data, where the right to use a voice must be documented. If you are comparing self-serve tools as well, our guide to AI data annotation platforms covers software options.
Conclusion
The best speech data partner depends on where your model is weakest. Appen, LXT and TELUS Digital suit large multilingual programs that need many locales and long-running delivery. Defined.ai, DataOcean AI and Magic Data are strong starting points when a licensed corpus can replace months of collection, with Magic Data particularly useful for conversational Asian language data. Welo Data and Shaip offer detailed annotation menus that run from diarized transcription to emotion tagging and sound event labels, while Andovar and Sigma AI serve teams that need studio recordings, low-resource languages or careful paralinguistic labeling.
Run a short paid pilot with two or three vendors on the same audio and the same style guide, then compare word error rate on a held-out set, turnaround and the quality of speaker metadata. The vendor that best reproduces your production conditions, whether noisy calls, accented speech or real two-way dialogue, will usually deliver the largest accuracy gain per hour of data.
