Buying speech data is a procurement decision with unusually high downside risk: a dataset with sloppy transcripts, missing consent records, or the wrong dialect can quietly poison a model for months before anyone traces the regression back to the source. Yet most teams evaluating a speech data provider for the first time have no rubric beyond price per hour.
This guide gives you that rubric. It covers how to scope your requirements precisely, when to buy off-the-shelf versus commission custom collection, how to audit a vendor's sample in 30 minutes, what pricing actually looks like in the market, and what must be in the contract before you sign. If you are new to the underlying technology, start with our primer on what automatic speech recognition is and come back.
The short version: scope in writing before you talk to anyone, never buy without auditing a sample, and never sign without seeing consent documentation. Everything below expands on those three rules.
Scoping Your Speech Data Requirements
Vague requirements produce vague quotes. Before contacting any speech dataset vendors, write down the following and treat it as your requirements document:
- Language and dialect. "Arabic" is not a specification; Moroccan Darija and Gulf Arabic are mutually difficult. Name the variety, the region, and whether code-switching with another language (e.g., French, English) is acceptable or desired.
- Hours of audio. Fine-tuning a modern multilingual ASR model on a new domain typically needs tens to a few hundred hours; training-heavy use cases need more. Decide your number and whether it counts total audio or speech-only (silence-trimmed) hours. Vendors differ, and the difference can be 20-30%.
- Speaker count and diversity. More speakers at fewer minutes each generalizes better than a few speakers at many hours. Specify minimum unique speakers, gender balance, and age distribution.
- Domain and speech style. Conversational telephone-style dialogue, scripted prompts, call-center scenarios, and read speech behave very differently in training. Match the data to your production use case.
- Audio specifications. Sample rate (8 kHz telephony vs. 16 kHz+ wideband), mono vs. dual-channel (one speaker per channel, essential for diarization and clean per-speaker transcripts), codec, and acceptable background-noise conditions.
- Transcript format. Verbatim vs. cleaned, time alignment granularity (utterance-level vs. word-level), speaker labels, tagging of non-speech events, and file format (JSON, TextGrid, CTM, plain text).
Our speech data collection guide goes deeper on each of these parameters and how they affect model performance.
Off-the-Shelf vs. Custom Collection: A Decision Framework
The first fork in the road is whether an existing dataset can serve your need or whether you should commission collection.
| Factor | Off-the-shelf favors you when… | Custom collection favors you when… |
|---|---|---|
| Timeline | You need data this month | You can wait 6-16 weeks |
| Budget | You want the lowest $/hr | You can pay a premium for fit |
| Language/dialect | A catalog dataset matches your variety | Your dialect or domain isn't on the market |
| Domain | General conversational data is adequate | You need scripted scenarios, specific vocabulary, or your own prompts |
| Recording conditions | Standard conditions are fine | You need specific devices, environments, or channels |
| Exclusivity | Shared licensing is acceptable | Competitors training on the same data is a problem |
| Iteration | One-off purchase | You expect follow-on batches and spec changes |
In practice many buyers do both: license 500-1,000 hours off-the-shelf from a dataset catalog to establish baseline performance, then commission a smaller custom speech data collection run to cover the gaps the baseline reveals. That sequencing is usually cheaper than guessing your full custom spec upfront.
How to Evaluate Speech Dataset Vendors
Any credible vendor will provide a sample on request, typically 10-30 minutes of audio with matching transcripts and metadata. If a vendor refuses to provide samples, stop; there is nothing to discuss.
The 30-minute sample audit
You do not need a research team to audit a sample. One person with headphones and a text editor can do it in half an hour:
- Audio quality (10 minutes). Open several files in an audio editor. Check the sample rate and channel layout match the spec sheet. On dual-channel recordings, confirm each speaker is genuinely isolated on their own channel with minimal bleed. Listen for clipping, dropouts, aggressive compression artifacts, and whether background noise levels are consistent with the stated recording conditions. Play files from different speakers to confirm they are actually different people.
- Transcript accuracy (10 minutes). Pick three files at random and read the transcript while listening. Count errors per minute of audio: substituted words, missing hesitations (if verbatim was promised), wrong speaker labels, misaligned timestamps. A well-QA'd conversational transcript should feel nearly frictionless to follow. If you are stumbling every few sentences, the full corpus will be worse; samples are usually a vendor's best foot forward.
- Metadata completeness (5 minutes). For every speaker in the sample, confirm the promised fields are present and plausible: speaker ID, gender, age band, region/dialect, recording device or channel. Check that speaker IDs are consistent across files.
- Speaker diversity (5 minutes). Tabulate the sample's speakers by gender, age band, and region against the claimed distribution of the full dataset. A sample drawn honestly should roughly reflect it.
Consent documentation
Ask to see a redacted example of the consent form speakers signed, and ask how consent records map to audio files. You want per-speaker, documented consent that explicitly covers commercial use and AI/ML training, not a generic terms-of-service checkbox. This matters for GDPR compliance and, increasingly, for your own downstream customers' vendor audits. If the vendor hedges here, walk away regardless of price.
QA evidence
Ask three questions: What fraction of transcripts get a second human pass? What word error threshold triggers re-transcription? Can you share your QA rubric? Vendors with a real quality process answer these immediately and specifically. "Everything is human-checked" with no numbers behind it is marketing, not QA.
Speech Data Pricing Benchmarks
Market pricing for transcribed conversational speech clusters into recognizable bands:
| Product type | Typical range ($/hr of audio) | What drives the price |
|---|---|---|
| Open/academic corpora | Free | Licensing usually blocks or complicates commercial use |
| Off-the-shelf, major languages | $50-100 | Volume, transcription depth, license breadth |
| Off-the-shelf, low-resource languages | $60-150 | Speaker recruitment difficulty, scarce transcription talent |
| Custom collection | $150-400+ | Demographic quotas, scripted scenarios, exclusivity, timeline |
SpeechData.ai's catalog datasets sit at $60-95 per hour across 60 languages, with 500-2,000 hours and 50-200 native speakers per dataset. When comparing quotes, normalize carefully: confirm whether the price is per audio hour or per speech hour, whether transcripts and metadata are included or itemized, and what license scope the price buys. A $55/hr quote with evaluation-only licensing is more expensive than an $85/hr quote with a full commercial training license. Our guide on how to buy AI training data covers cross-vendor price normalization in more detail.
Procurement Process and Timeline
A realistic off-the-shelf purchase runs four to eight weeks end to end:
- Week 1: Scoping. Finalize your requirements document. Shortlist 3-5 vendors.
- Weeks 1-2: Samples and audit. Request samples under NDA if needed. Run the 30-minute audit on each. Eliminate anyone who fails audio, transcript, or consent checks.
- Weeks 2-3: Quotes and normalization. Get written quotes against your exact spec. Normalize pricing and license scope side by side.
- Weeks 3-6: Legal and security review. Data license agreement review, data processing terms, and (for enterprise buyers) vendor security questionnaires. This is the usual bottleneck; start legal review before you have picked a final vendor if your timeline is tight.
- Weeks 4-8: Delivery and acceptance. Insist on an acceptance window (10-15 business days) during which you validate the full delivery against the sample-audit criteria before final payment.
Custom collection adds 6-16 weeks of recruitment, recording, and transcription on top, which is why the off-the-shelf-first sequencing above is so common.
Contract Essentials
The license agreement matters as much as the data. At minimum, nail down:
- License scope: internal model training and evaluation, including derivative models and their commercial deployment. Confirm model weights are yours.
- Redistribution and sublicensing: usually excluded, so make sure your actual deployment pattern (e.g., shipping models to customers) doesn't accidentally require it.
- Warranties: that the vendor has the rights to license the data and that speaker consent covers your use.
- Indemnification: vendor indemnity for third-party IP and privacy claims arising from the data itself.
- Acceptance criteria: objective quality thresholds tied to payment, referencing your audit rubric.
- Deletion and audit terms: what happens to the data if the agreement terminates, and what consent evidence the vendor must retain and produce.
We cover every clause in depth in our companion guide, Speech Data Licensing Explained.
The Full Evaluation Checklist
Use this table as your scorecard when comparing speech dataset vendors:
| # | Criterion | What "good" looks like | Pass/Fail |
|---|---|---|---|
| 1 | Sample availability | Sample provided within days, with transcripts and metadata | |
| 2 | Audio spec compliance | Sample rate, channels, and format match the spec sheet exactly | |
| 3 | Dual-channel separation | One speaker per channel, minimal bleed | |
| 4 | Transcript accuracy | Near-frictionless read-along; errors rare and minor | |
| 5 | Time alignment | Timestamps accurate at the promised granularity | |
| 6 | Metadata completeness | Every speaker has ID, gender, age band, region, device | |
| 7 | Speaker diversity | Sample distribution matches claimed dataset distribution | |
| 8 | Consent chain | Per-speaker consent covering commercial AI training, mapped to files | |
| 9 | QA evidence | Documented process with named thresholds and review rates | |
| 10 | License scope | Commercial training license; you own resulting model weights | |
| 11 | Warranties & indemnity | Rights and consent warranties; IP/privacy indemnification | |
| 12 | Price normalization | $/hr comparable after adjusting for license and inclusions | |
| 13 | Acceptance window | 10+ business days to validate full delivery | |
| 14 | Follow-on capability | Vendor can do custom collection if gaps appear later |
A vendor should pass items 1-11 outright; 12-14 are negotiation and planning items. Two or more hard failures in the first eleven means move on. The market has enough credible speech data providers that you never need to compromise on consent or quality.
Talk to us
SpeechData.ai offers off-the-shelf conversational speech datasets in 60 languages, with dual-channel audio, time-aligned transcripts, full speaker metadata, and a documented consent chain, at $60-95 per hour with samples available on request. Browse the dataset catalog or contact us to request a sample against your spec.