Speech Data Collection Tools Compared: Common Voice Has Just 16 Validated Hours of Hindi

Common Voice, the largest public-domain speech corpus, holds 29,295 validated hours across 295 languages. Only 15.7 of those hours are Hindi. If you need speech data for AI training in an Indian language, the open data will not carry you, and you will end up choosing a speech data collection tool to gather your own.

The hard part is not recording. It is finding speakers, proving consent and checking every clip. We read the release notes, the papers and the vendors' own pages, and compared five ways to collect an audio dataset on the jobs that actually decide the outcome.

What a speech data collection tool has to do

A speech dataset is pairs of audio and text, plus who spoke and under what terms. A tool that only records audio does a fifth of the job. The rest is recruiting speakers, recording consent, logging demographics and checking that each clip says what the text says.

We judge every option below on five questions. Who records? How is consent captured? Which languages are realistic? Who checks quality? And how do you pay? A tool that answers three of these well and two badly can still be the right choice, as long as you know which two.

Scripted speech and spontaneous speech need different tools

Scripted speech means a speaker reads a sentence you wrote. The transcript exists before the audio, so checking is a yes or no question: did they say it? Spontaneous speech means a speaker answers a prompt in their own words. Somebody then has to transcribe it, which costs far more.

Most tools handle scripted reading well. Spontaneous speech is where field operations and vendors earn their fees. We cover the method side of this choice in how to build a speech dataset for ASR.

Speech data collection tools at a glance

All vendor facts below come from each company's own pages as of October 2026. Where a page did not state something, we say so instead of guessing.

The table hides one asymmetry. The first four bring speakers. The last two assume you already have them. That single question, whether you need people or only a tool, settles most decisions before price comes up.

Mozilla Common Voice: the open baseline

Common Voice is a volunteer project. Per the Common Voice paper, contributors record by reading sentences shown on screen, and other contributors vote on each clip. A clip that first gets two up-votes is marked valid; two down-votes mark it invalidArdila et al., Common Voice: A Massively-Multilingual Speech Corpus, LREC 2020, arXiv 1912.06670. Describes the validation rule and the CC0 release.. The corpus is released under CC0, so you can train commercial models on it.

That makes it the first place to look and the cheapest data you will ever get. It also means you get what volunteers chose to record, not what your model needs. You cannot ask Common Voice for 200 speakers from rural Bihar reading banking commands.

Validation is the bottleneck, not recording

The release statistics in the cv-dataset repository show the gap between audio recorded and audio checked. For Indian languages, the gap is wide.

Across all languages, about 69% of recorded hours are validatedCommon Voice Scripted Speech v27.0, 11 September 2026: 42,593 total hours, 29,295 validated. Our division.. For Bengali it is about 4%. We think this is the most useful number in the whole comparison. Any crowdsourced project lives or dies on the second listener, and you should budget for listening time before recording time.

Bhasha Daan, the government volunteer channel

India's Bhashini mission runs a similar volunteer scheme called Bhasha Daan. A 2022 government factsheet describes its Bolo India strand as donating your voice by recording sentences and validating others' recordings. Contributions can be anonymous. The factsheet is from August 2022, so treat it as a description of the model, not of current volumes.

Karya and the field-collection model

Karya started as an open-source crowdsourcing platform from Microsoft Research and now runs as a data company. Its homepage describes "large-scale conversational datasets across 22 official Indian languages" and custom data work. It does not publish prices or pay rates.

Press reporting fills part of the gap. A 2023 Thomson Reuters Foundation report said Karya pays workers about $5 an hour and that workers own part of the data they generateThomson Reuters Foundation, Context, 4 December 2023. Reported company statement, not independently audited. Tier 2 source.. That is the company's own claim as reported, and it is three years old.

What IndicVoices shows about scale

The best public account of Indian field collection is the IndicVoices paper from AI4Bharat and IIT Madras. The team used the open-source Karya Android app because it works offline and syncs later. That matters in districts with patchy connectivity.

The numbers show what "field operation" means. IndicVoices collected 7,348 hours from 16,237 speakers in 145 districts, with 1,893 people working on itJaved et al., IndicVoices, arXiv 2403.01926v1, March 2024. Section 1 and Section 4.. Each district had 1 to 2 local mobilisers and 5 to 7 coordinators. A coordinator sat with each participant, checked every response on the spot and asked for repeats.

Even then, the paper reports problems a tool cannot fix. Some registration forms carried false age or gender to meet quotas. Some participant IDs had several different voices behind them. The team added a short identity video, reviewed only by female QC staff and deleted after checking. We would not try to reach this scale without a partner that already has people on the ground.

Data vendors: Appen and Defined.ai

If you need a contract, a delivery date and someone else's legal team, a vendor is the honest answer. Both companies offer two things: licences for datasets they already hold, and custom collection.

Appen's datasets page lists 597 off-the-shelf datasets across nine categories and cites "500+ global locales". It says custom collection suits projects needing "specific demographics, acoustic environments, specialist domains, or a controlled collection protocol". On consent, it states that all data is collected "under explicit contributor consent with documented provenance". Pricing is by quote.

Defined.ai's data collection page cites 1.6M+ contributors in 150+ countries and 500+ languages. It states that every dataset is "fully consented, copyright-cleared, and privacy-compliant". For audio it lists conversational dialogues, IVR interactions and studio recordings. It does not publish a price model.

Where vendors are the better choice

Vendors win when your speakers are hard to find and you have budget but no time. They also win when procurement needs a named counterparty for consent and provenance. A web form cannot recruit 500 native Somali speakers for you. That is the honest limit of the DIY options below.

Collecting speech data yourself with a web form

The DIY route fits a different situation. You already have speakers: your own users, a partner NGO's field staff, students, or a panel you recruit. You need them to read a script and send clean clips, and you need it this week.

Google Forms needs a separate recorder

Google Forms has no voice question. Google's help page lists file upload as the closest type, and states that responders must sign in to a Google Account to answer it. So each speaker records in another app, finds the file, and uploads it. For a 50-sentence script, that is 50 round trips. We go through the workarounds in whether Google Forms can record audio.

A voice form records in the phone browser

Would You Rather is our product, so read this section with that in mind. Each script sentence becomes one voice question. The speaker sees one sentence at a time, taps record, listens back, re-records if needed and moves on. No app install, and anonymous answers are a per-form setting.

For script-reading projects we can switch on a transcript mode. The script sentence is prefilled as an editable transcript under the recorder, and the speaker corrects it only if they said something different. Typing in English letters offers suggestions in the target Indian script. Each voice question also takes a minimum and a maximum length, so a 1-second accidental tap is rejected. Transcript mode is not yet a switch in the builder, so ask us to set it up.

Recordings land next to the speaker's other answers, with audio links in a live Google Sheet or CSV. What it does not do matters as much. There is no built-in second-listener vote, no auto-transcription and no speaker pool. You bring the people and the QC. We cover the general recording flow in collecting voice recordings online.

How to choose a speech data collection tool

Decide in this order. Each step removes options, and most teams are down to one before they reach step 4.

When another option beats a web form

Where this comparison is weakest

We did not run a paid project with any vendor, so we cannot compare their real turnaround or reject rates. Everything about Appen, Defined.ai and Karya comes from their own pages, plus one 2023 press report for Karya's pay. Vendor pages describe intent, not audited outcomes.

The Common Voice numbers are exact but narrow. They count hours, not quality. A validated hour from 3 speakers is worth far less for training than one from 300, and the headline totals hide that. Check speaker counts per language before you trust an hour figure.

IndicVoices is one project, run by a well-funded academic team. Its processes describe what worked for them, not a minimum. And our own product's section is written by us. The most important line in it is the list of things it does not do.

Where to start this week

Open the Common Voice release statistics for your language and write down three numbers: validated hours, speakers, and the share of female speakers. If any is too low, you have your case for collecting.

Then write 50 script sentences, put them in a voice form and send it to 10 speakers you already know. Listen to all 500 clips yourself. You will learn more about your script, your instructions and your reject reasons in that afternoon than from any vendor call. If your speakers write in Indian scripts, read collecting answers in Hindi, Tamil or any language first.

References

1. Ardila, R. et al., Common Voice: A Massively-Multilingual Speech Corpus, LREC 2020. Used for the recording and validation process and the CC0 licence.
2. Mozilla Common Voice, cv-dataset repository, release statistics for Scripted Speech v27.0, 11 September 2026. Used for total, validated and per-language hours.
3. Javed, T. et al., IndicVoices: Towards building an Inclusive Multilingual Speech Dataset for Indian Languages, arXiv preprint, March 2024. Used for field collection scale, staffing, the Karya app and QC problems.
4. Ministry of Electronics and IT, BHASHINI factsheet, Press Information Bureau, 26 August 2022. Used for the Bhasha Daan description.
5. Karya, homepage, accessed 8 October 2026; and Thomson Reuters Foundation, Context, "India turns to AI to capture its 121 languages", 4 December 2023. Used for Karya's stated offering and reported pay.
6. Appen, off-the-shelf datasets page, accessed 8 October 2026. Used for catalogue size, locales, consent statement and custom collection.
7. Defined.ai, data collection page, accessed 8 October 2026. Used for contributor numbers, languages and consent statement.
8. Google, Choose a type of question for your form, accessed 8 October 2026. Used for the file upload sign-in rule.

The weakest point in this source base: every vendor fact is self-reported marketing copy, and the only figure on Karya's pay comes from a 2023 press report of a company statement.