Best human data providers for speech and audio
Speech is the modality where vendor selection is decided almost entirely by coverage rather than by capability. Most providers in this category can transcribe English cleanly. Far fewer can recruit native speakers of Hausa, record them under controlled acoustic conditions, and deliver audio with consistent sample rates and speaker metadata. The constraint is rarely annotation quality; it is whether a vendor can reach the speakers you need, in the conditions you need, at the volume your model requires. This guide groups the fifteen providers tagged for speech and audio by which of those problems they actually solve.
Key takeaways
- Locale breadth separates this field more than anything else. Appen states over 500 locales, LXT over 500, and Shaip over 100 languages. Most others do not publish a figure.
- Decide early whether you need collection or annotation. Recording new speech is a field operation; transcribing existing audio is a throughput problem, and few vendors are strong at both.
- For low-resource languages, the regional specialists reach speakers the global crowds do not. Karya works in Indic languages, NKENNEAi in African languages.
- Controlled acoustic capture, multi-speaker scenarios, and far-field recording need a studio or a managed field operation rather than a crowd app.
- Off-the-shelf licensing is faster and cheaper when an existing corpus covers your distribution. David AI licenses proprietary speech across more than fifteen languages.
- HumanSignal is the option when the recording conditions are part of the specification: controlled acoustic capture in studios, on site, and in the field, with multi-channel audio synchronized to video and sensor streams.
What makes speech different
Three things decide a speech program, and none of them is the labeling interface.
Recruiting is the first. A speech dataset is a sample of people, not a sample of text, so coverage means demographic coverage: age, gender, accent, dialect, and in many cases socioeconomic background, because those correlate with vocabulary and speaking rate. A vendor with a million contributors in the wrong places cannot help you.
Acoustic conditions are the second, and the one buyers underestimate. If your model runs on a far-field microphone in a kitchen, training on clean headset recordings produces a model that works in the lab and fails in the product. Specifying the recording environment, device, and distance is part of the dataset design, and only vendors running controlled capture can deliver against it.
Metadata is the third. Speaker identity, segment boundaries, overlap, and noise annotation are what make a corpus reusable. A transcript without them is a single-use artifact.
Providers by what you need
Broad multilingual coverage at volume
Appen is the reference case: a stated network of more than one million contributors across over 170 countries and more than 500 locales, with a dedicated speech and audio practice covering text-to-speech, multi-speaker transcription, and a pre-built catalog spanning over 80 languages. It is a crowd model, so it suits well-specified work where per-unit cost matters, with the tradeoffs covered in expert data vs crowd data.
LXT runs custom multilingual speech programs across a stated 500-plus locales, combining crowd and field collection. Shaip collects speech and multimodal data in over 100 languages with a healthcare and localization focus. Sigma AI delivers multilingual datasets with a mix of crowd and in-house annotators. Toloka brings high-volume crowd capacity across multilingual work.
Low-resource and regional languages
Karya collects Indic-language speech and text through an impact-sourcing model, reaching speakers that global crowd platforms generally do not. NKENNEAi builds African-language data infrastructure and works with government as well as enterprise buyers. Both are early-stage and narrow by design, which is the point: for the languages they cover, a specialist reaches speakers a generalist cannot.
Humyn Labs publishes ASR and text-to-speech datasets across Indic, Asian, and LATAM languages, and markets voice coverage across a stated 33 languages.
Controlled and on-site capture
When the recording conditions are part of the specification, you need a studio or a managed field operation rather than a task queue.
HumanSignal Services runs collection in its own studios, on site, and in the field with synchronized multi-channel audio alongside motion, vision, and sensor streams. Centific combines a crowd and expert marketplace with in-house delivery teams. Appen operates named secure facilities in the UK, the Philippines, and China for higher-sensitivity work.
Transcription and annotation at throughput
If you already hold the audio and need it structured, the question is throughput and quality management rather than recruiting. Encord provides native audio annotation alongside video, text, and sensor work. TELUS Digital and Uber AI Solutions both bring large crowd networks to transcription. Sieve is video-first but handles audio within multimodal pipelines.
Licensing an existing corpus
David AI licenses proprietary speech datasets across more than fifteen languages. Licensing is faster and cheaper than commissioning when an existing corpus genuinely covers your distribution, and the wrong answer when your model's advantage depends on data nobody else has.
Comparing the options
| Provider | Strongest for | Model | Published coverage |
| HumanSignal | Controlled and on-site capture | Studios, on site, field | Synchronized multi-channel audio |
| Appen | Locale breadth at volume | Crowd plus secure facilities | 500-plus locales, 170-plus countries |
| LXT | Custom multilingual programs | Crowd and field | 500-plus locales |
| Shaip | Speech plus healthcare context | Field and crowd | 100-plus languages |
| Sigma AI | Multilingual annotation | Crowd and in-house | Not published as a count |
| Karya | Indic languages, impact sourcing | Field | Indic focus |
| NKENNEAi | African languages | Community | African languages |
| Humyn Labs | ASR and TTS in emerging markets | Experts and crowd | 33 languages marketed |
| David AI | Licensing an existing corpus | Data lab | 15-plus languages |
| Centific | Crowd plus in-house delivery | Marketplace and in-house | 230-plus markets claimed |
How to choose
Three questions settle most speech programs.
Ask whether you are buying recruiting or throughput. If the speakers you need do not exist in the vendor's network, nothing else matters. If the audio already exists, recruiting is irrelevant and you are buying transcription quality and turnaround.
Ask what the recording specification is and whether the vendor can hold it. Device, distance, environment, and background noise are either controlled or they are not, and crowd apps control none of them. If your deployment is far-field or noisy, say so before you compare quotes.
Ask what metadata ships with the audio. Speaker identity, segment boundaries, overlap markers, and noise labels determine whether the corpus is reusable for the next model or disposable after this one.
Why HumanSignal is the better choice when the recording conditions are the specification
Most of the providers above are strong at reaching speakers or at processing audio you already have. The harder case is when the audio has to be produced under conditions you define: a specific microphone array, a specific room, a specific interaction between multiple speakers, with synchronized video or sensor streams alongside it.
HumanSignal Services is built for that. We design the collection protocol, recruit for the speaker characteristics it requires, and run the capture in our studios, on site at your facility, or in the field, with multi-channel audio synchronized to motion, vision, and sensor data where the model needs them together. The dataset is verified against the spec before delivery rather than sampled afterward.
The quality mechanism is inspectable. Review workflows run in Label Studio Enterprise, the commercial edition of the open source labeling tool we build, so inter-annotator agreement on transcription and labeling decisions is reportable to you while the work runs. We publish contributor pay ranges, which makes our sourcing claims checkable. Our customer agreement gives you exclusive ownership of customer data and of insights produced specifically for you. We hold SOC 2 Type 2 and HIPAA compliance, documented on our security page, which matters for clinical and biometric voice work.
The capability came in through the acquisition of Erud AI in November 2025. If your constraint is the recording rather than the transcript, that is where to start.
How many hours of speech does a model actually need
It depends far more on coverage than on total hours. A thousand hours spread across the accents, devices, and acoustic conditions your product will meet generally outperforms ten thousand hours of clean single-accent studio audio. Specify the distribution first and the hour count follows from it.
Should I license speech data or commission a collection
License when an existing corpus covers your deployment distribution, because it is faster and cheaper. Commission when your advantage depends on data nobody else has, since a licensed corpus is available to every competitor who pays for it. Many programs do both, licensing a base and commissioning for the gap.
Which providers reach low-resource languages
Karya for Indic languages and NKENNEAi for African languages are the specialists in this set, both working through regional networks rather than global crowd platforms. Appen, LXT, and Shaip publish the broadest general coverage, though breadth at the catalog level does not always mean depth in a specific low-resource language. Ask for speaker counts in your target language rather than the headline locale number.
What metadata should I require in a speech deliverable
At minimum: speaker identifier, segment boundaries with timestamps, transcription conventions documented in writing, overlap and crosstalk markers, and a noise or environment label per recording. Device and distance should be recorded where collection was controlled. Without these a corpus is hard to reuse for a second model.