7 best speech data providers in 2026: buyers and owners

Put the requirement in writing before asking for quotes. These five checks apply both to a buyer comparing offers and to an owner describing an archive. Inspect samples only for the purpose their permission allows.

1. Match the speakers and speech to the task

Specify the regional accents, vocabulary and speech style you need. Scripted commands and spontaneous conversation answer different requirements. Ask how many distinct speakers and how much audio cover each required combination, including what is missing. An archive owner should mark unknown coverage rather than infer it from a language label.

Test: Ask the provider to identify the offered recordings matching your accent and vocabulary requirements. Compare that breakdown with an authorized sample. “We list that language, so all its regional accents are covered” is an illustrative red flag.

2. Match the recording conditions to the setting

State whether the task needs studio speech, calls, distant microphones, background sound or overlapping speakers. Request the original device and format, channel layout and processing history. Two channels do not necessarily mean two separately recorded speakers. Noise reduction or channel mixing belongs in that history, too.

Test: Follow one sample from its source recording to the delivered file. “Everything is WAV, so the original device does not matter” leaves your question unanswered. Use the audio review guide to record the technical evidence.

3. Check what the transcript keeps and how it follows the audio

Agree whether hesitations, false starts, uncertain words and simultaneous speech are retained or marked. Then check which audio version the transcript describes. A timestamp does not establish word-level alignment; ask what each boundary means and how speaker changes are represented.

Test: Follow one passage from recording to transcript, including an interruption or uncertain word. “Your team can reconstruct the timing after delivery” signals extra work to price. Owners should identify transcripts made from earlier edits. The transcript and provenance guide separates these records.

4. Confirm the permitted uses and the owner's role

Establish whether the offer licenses existing recordings, commissions new ones, or pays for preparation. Permission to hear a sample does not settle permission to train a model, generate a voice or redistribute recordings. Ask which records support each intended use and who resolves a gap. Owners need to understand retained rights and future-use approval as well.

Test: Request the license scope and its referenced permission records. “The speakers agreed to be recorded, so every later AI use is included” is an illustrative red flag, not a legal conclusion about any provider.

5. Compare the deliverable and the full commercial terms

Define the billable unit: raw recording hours, transcribed hours or material that passes agreed checks. Separate license access, collection, transcription, review, corrections and delivery. Ask who decides acceptance and who pays for rejected material to be corrected. Owners need their proceeds, deductions and payment trigger stated separately from the buyer's invoice.

Test: Have each provider price the same scope and identify what could increase the bill. “Pay for uploaded hours first; acceptance comes later” leaves a material commitment undefined. A clear custom quote is more useful than an hourly price with no deliverable attached.

Use this table to prepare a written quote request. Buyers need the exact recordings, permitted uses, full price, acceptance checks and remedies; owners need their proceeds, preparation costs and payment trigger. “Not published” means the field is not stated in the cited public offer or terms. Confirm it in the selected agreement. Source dates and their basis appear in Sources.

Rightmark: best for licensing an existing speech archive

We help owners license private recordings to AI companies and help buyers source audio against a specification. Our audio route covers business calls, multilingual speech and podcast archives. The owner brings an existing asset; the buyer brings a requirement for speakers, language, task and format. Our buyer workflow connects those two sides through a proposed scope and representative scrubbed samples under NDA. Buyers can evaluate the material before taking a license, while owners can judge the proposed use before authorizing it. Evaluation access does not itself permit training.

The fee arrangements make the owner's proceeds and the buyer's cost separately visible. On standard licenses, owners keep the agreed owner payout and buyers pay our fee on top. For premium and bespoke deals, owners keep 75% to 85%, agreed in writing before a sale. An owner can therefore compare the proposed proceeds with the work needed to prepare the archive; a buyer can distinguish the data payment from our fee. Keeping the agreed payout does not mean receiving the buyer's entire spend. Our Trust page says, “Nothing is licensed until you sign the agreement and approve its scope.”

Appraisals don't prove you can license recordings, as our preparation note explains. Transcripts accompany audio where they already exist, so an untranscribed collection leaves preparation to resolve. Our public speech coverage is less detailed than Shaip's catalog of speech types and recording metadata. A buyer choosing a specific accent needs the actual proposed recordings, not a broad audio category. We currently exclude health data and data about children. Check permission records and review permitted samples before pricing those recordings.

Best for owners licensing an existing speech archive; buyers needing published speech-type and recording detail should also compare Shaip's catalog.

Defined.ai: best for choosing catalog speech by recording source

Defined.ai's audio catalog separates on-device scripted monologues, simulated call-center dialogue and live podcasts. Those are materially different starting points. Scripted speech suits a requirement for specified utterances; a simulated call supplies dialogue without being a recording of an actual customer interaction. Podcasts supply long-form speech recorded by podcasters. The catalog describes sample rate, bit depth and channel structure, alongside locale, domain and volume metadata. It also identifies human transcription on relevant datasets and transcription services for longer sources. A podcast archive and a transcribed command collection can therefore leave very different amounts of preparation in the buyer's budget.

For missing material, its speech services offer custom voice collection with phonetically balanced scripts. A TTS project can specify the voice and sounds to record instead of adapting an unrelated conversation corpus. Defined.ai also offers human assessments of pronunciation, naturalness and context, through an API or a managed service. The service belongs after a voice produces speech: listeners judge pronunciation and how natural it sounds in context. That is a different purchase from recordings for training. A TTS team can commission its voice and script, then separately evaluate the speech its model produces. Its partner program reviews owners' data, metadata and licensing rights. Owners license existing archives; service suppliers take collection or annotation assignments. An unnamed podcast and video media aggregator says there, “In a shifting media landscape, finding new revenue streams isn’t a bonus but a necessity.”

The license terms allow commercial model use without exposing recordings or recognizable contributor content. Raw-audio redistribution needs written authorization.

Best for choosing catalog speech by recording source; raw-audio distributors need express written permission.

Appen: best for catalog speech plus targeted collection

Appen's catalog supports a practical split: license a baseline, then commission missing demographics, acoustic settings or domain material. Its dataset specification records source, consent, annotation and licensing information, making that split a comparison of documented material rather than headline hours. Appen's speech services include transcription, speaker diarization, acoustic-event labels and tone or emotion annotation. A transcript records the words; speaker attribution connects them to the people speaking, while event and tone labels describe other parts of the recording. The right combination depends on whether the task concerns the spoken content, the exchange between speakers or the surrounding sound.

Microsoft Translator's difficulty was sourcing and annotating language data from smaller speaker communities. In Appen's case listed in October 2024, local native speakers supplied samples that were transcribed and translated. The work also involved language structure and cultural context, including multiple translations for gender-ambiguous sentences. Marco Casalaina, identified as VP of Products for Microsoft Azure AI, cites Appen's “ability to source speakers, even for under-resourced languages”. That makes the case relevant when the obstacle is finding suitable speakers and preparing their contributions. Buying more hours of a well-represented language would not solve that problem; the missing language community has to be part of the collection brief.

For qualifying owners, Appen's enterprise-data program is a separate licensing relationship. Owners retain ownership and approve scope and samples. Appen covers extraction, anonymization, preparation and transfer; compensation is agreed before extraction and paid on acceptance, not submission. Owners needing an advance must agree that separately. Speech-archive eligibility still needs confirmation.

Best for catalog speech plus targeted collection; enterprise owners needing an advance require terms beyond the published acceptance-based payment.

Shaip: best for speech datasets with transcripts and recording metadata

Shaip sells pre-built speech datasets and provides custom collection and annotation. Its speech catalog pairs recordings with transcripts, speaker demographics, recording-environment tags and timestamps. That makes it useful when the purchase needs to include the text and the context of the audio. Conversation, scripted monologue and TTS recordings remain different products: a buyer seeking conversational speech should select that type before narrowing by country or sample rate. The accompanying fields help describe the selection, although a timestamp field alone does not establish word-level alignment.

Its collection service selects speakers against requirements such as dialect and language, with studio recording and voice collection in noisy environments. The setting belongs in the commission: clean studio speech cannot stand in for the sound of a voice competing with background noise. Its annotation service works to client-specific guidelines, with pricing shaped by volume, complexity and customization. A buyer can use those guidelines to define what to retain in the transcript and how to mark uncertain speech, then make the agreed convention part of acceptance. An unnamed Senior Applied Scientist at Oracle gives general service feedback on Shaip's homepage: “I want to express my appreciation for the support and professionalism your team has consistently provided.” For an owner commissioning annotation, that preparation has a cost separate from any future license proceeds.

A refundable service pilot requires an express right in the services agreement or statement of work; otherwise Shaip's payment terms make payments final and non-refundable. Keep that distinct from the correction process. A requirement to fix agreed transcript errors and a right to recover the fee serve different purposes, and the contract should settle which applies to your evaluation.

Best for speech datasets with transcripts and recording metadata; a refundable service pilot needs an express right in the Shaip contract.

Protege: best for curated conversational audio and paired transcripts

Protege's audio offer starts with real conversational recordings and curates them around linguistic, acoustic and contextual requirements. Human-reviewed transcripts are paired with the audio. This suits a buyer whose requirement is a particular mixture of conversations, rather than a script to be recorded from scratch. Its iterative selection process allows the proposed material to be refined around that requirement. The distinction matters when a broad library contains useful speech but only some of it belongs in the purchase. The transcript and recording need to travel together through that selection.

Protege lists echo, reverb, clicks and drops among its audio checks. Those findings help distinguish a recording's useful acoustic conditions from defects that need attention; the relevant result is the condition of the selected files. For owners, the audio partnership route reviews the library and rights framework, structures revenue sharing and brings material into qualified deals. Owners retain underlying ownership while licensing defined uses. That offers a way to earn from an archive without selling the archive itself. A cultural ambition also appears in its December 2025 CBS announcement: David Ulmer, identified as Cambodian Broadcasting Service's CEO, says, “We envision a world where future AI models have a full understanding of Cambodian culture, text, and speech.” His statement describes a hoped-for use of the material.

Protege's published media policy excludes copying a performer's name, image or likeness, including voice. A team trying to reproduce an identifiable performer's voice therefore needs a different permission. This does not exclude every synthetic-voice project. Its platform terms restrict training to the statement of work's authorized purposes, with a conflicting signed statement of work controlling. The speech agreement must establish the permitted use; the public policy supplies no automatic exception.

Best for curated conversational audio; teams needing to copy a performer's voice should choose an arrangement that expressly permits it.

iMerit, part of EXL: best for turning existing recordings into speech ground truth

iMerit works on audio a client already holds, producing transcripts and project-specific labels. Its transcription service uses human listeners, handles recordings with background noise and calibrates quality and throughput requirements with the client. It can add entity, intent and sentiment labels, with separate production and QA stages. That is useful when collecting more audio would leave the real problem untouched: the team still needs a reliable reference for the recordings it has. Archive owners using this service are purchasing preparation. EXL announced completion of its iMerit acquisition on August 3, 2026.

The Digital Green case shows why subject knowledge matters in transcription. Farmer.Chat received voice queries with local pronunciation, agricultural terminology, pauses and background noise, recorded on low-cost devices. According to iMerit's account, native-language experts transcribed the queries, with annotation and quality review organized on ANGO. The resulting ground truth was a human reference against which model output could be compared. Vineet Singh, identified as Digital Green's CTO, says, “The ground truth data enables us to benchmark models, understand errors, and design targeted strategies to improve performance.” The practical value was knowing what the farmers had actually said. A model error involving a crop name needs that reference before a team can decide what to improve.

Human listening also sets a clear boundary. If outside annotators may not hear your recordings, this described service is unsuitable without an authorized arrangement. For owners who can share the audio for transcription, the next decision is which labels make the archive useful to its intended buyer; transcription access and later model-training rights remain separate permissions.

Best for turning existing recordings into speech ground truth; teams prohibiting external listeners should use an authorized internal workflow.

DataForce by TransPerfect: best for commissioned speech and TTS recordings

DataForce, TransPerfect's AI data solutions division, recruits contributors for specified recording projects. Its collection service offers mobile, supervised, remote and studio setups, with participant opt-in and data-use agreements. The setup is part of the product: a mobile recording and a studio voice session answer different acoustic requirements. Buyers can commission the recording conditions they need instead of searching only among existing archives. Its transcription service adds client-specific tags for noise, incomplete words and unintelligible speech, plus speaker-turn or event timestamps where required. Those tags retain information a polished reading transcript would lose. Owners commissioning this work are paying to prepare audio, separate from any archive licensing deal.

During pandemic studio restrictions, iGenius needed a customized voice for Crystal. The published account describes a project with DataForce and MediaNEXT that brought voice talent, script design and remote recording together. The team agreed script length and sentence duration and planned a balance of English speech sounds; StudioNEXT supported recording remotely. Marco Bocchio, identified as iGenius's Machine Learning and Data Science Team Lead, describes “phonetically balanced sentences and corresponding high quality audio clips”. The useful decision came before recording: which voice would speak, which sounds the text would cover, and how long each sentence should be. A pile of unrelated recordings would leave those requirements unsettled.

For transcription, DataForce describes automatic checks and manual review of a percentage of the data, higher at the beginning and lower later. A buyer requiring a second human review of every utterance needs that coverage explicitly agreed. Sampled review and complete second review are different deliverables, with different work behind them. The scope should connect review coverage to corrections and the unit accepted for payment.

Best for commissioned speech and TTS recordings; full second human review needs an explicit commitment beyond the published sampled-review process.

Step 1: Narrow to two or three that fit the job

Choose two or three providers against one requirement: task, speakers, recording conditions, transcripts, rights, volume and budget. Record what is available and what still needs collection or preparation. Owners should also establish whether the conversation concerns licensing an archive or purchasing work on it. A contributor assignment is a third relationship.

Step 2: Ask every candidate the same seven questions

Keep the answer, supporting document, date, remaining gap and responsible person beside each question. These red-flag replies are illustrative, invented examples, not statements made by the providers.

1. “Which part of our requirement is available now, and which speakers, accents or speech types are missing?” Request a breakdown of the actual offered material. An owner should confirm whether existing archives are considered. Red flag: “That language is in our network, so your accents are covered.”
2. “Show a permitted sample with its device, channel arrangement and processing history.” Compare the evidence with your required setting. Red flag: “It is WAV, so it meets your noisy-conversation requirement.”
3. “Follow this transcript passage back to the exact audio version.” Check speaker changes, timing, overlap, uncertainty and edits. Red flag: “We edited for readability; you can reconstruct the missing timing.”
4. “Which documents authorize evaluation, training, voice generation and redistribution?” Identify permitted uses, exclusions and who resolves gaps. Owners should confirm retained rights and future-use approval. Red flag: “You may listen to it, so all AI uses are covered.”
5. “Who transcribes and reviews the material, how much is reviewed, and what proves acceptance?” Agree measures and corrections for this batch. Red flag: “Our company reports high accuracy; batch-level review details are unnecessary.”
6. “What is billable, what costs extra, and when is payment due?” Include minimums, rejected material and, for owners, proceeds and deductions. Red flag: “The hourly price tells you enough; we can settle acceptance later.”
7. “What bounded evaluation can we agree, and what happens if its requirements are missed?” Establish cost, allowed use and exit consequences. Red flag: “Start with the whole archive; we will settle evaluation limits afterward.”

Use the sample-pack checklist to organize the material for review, and keep the seven answers in a separate comparison sheet. Mark each point supported, needs clarification or unsuitable for your requirement, with a reason.

Step 3: Propose a bounded 30 to 60 day evaluation

Propose a paid evaluation where access or preparation has a cost. Agree a 30 to 60 day period, spending limit, named decision owner and written deliverable. Select the speaker and recording-condition slices to inspect, then specify files, transcript conventions, timing, review coverage and exclusions. A different minimum or evaluation format is a term to compare; this proposal does not mean every provider offers a trial.

Keep the work within the license. Inspection-only permission allows the agreed review, not model training. A training comparison needs explicit authorization and an agreed method. Record recipients, access, retention and what must be returned or deleted at the end.

Write the consequences of a missed requirement before starting. A wrong audio/transcript link, an absent required accent or missing agreed review should have a named decision owner and a correction or rejection path. State who pays for rework and what cancellation or refund provision applies. An owner should also cap preparation costs and identify the event triggering compensation. The evaluation can end with a smaller scope, a documented gap or no deal.

Choose the offer that supplies the recordings, preparation and permissions your task needs at a complete, understandable cost. Owners should make the same distinction between a licensing proposal and a preparation bill. A case study can justify a conversation; the actual material and agreement must justify the commitment.

Use the sample-pack checklist to prepare the evidence for your next provider conversation.