French speech dataset: 1,131 hours in stock
1,131 hours of unscripted French recorded in January and February 2026, mostly by speakers in Algeria, Madagascar, Morocco, Rwanda and Benin: 1,012 hours of two-person conversation, each speaker on their own channel, and 118 hours of single-speaker recordings, with machine transcripts and word-level timestamps. Download access within 24 hours of a signed order.
- In stock
- 1,131 hours
- Recordings
- 2,088
- Speakers
- 82+
- Audio
- 48 kHz WAV, stereo and mono
- Transcripts
- Machine, word timestamps
- Price
- $35 per hour, full set
Need speakers, topics or recording conditions this set does not cover? We also record speech data to your specification, quoted separately.
The video couldn’t start. Try playing it again or open the video directly.
Hear the French recordings
Unedited excerpts from the sample recordings we send to buyers, with the uncorrected machine transcript for the same stretch of audio.
Business role-play: a restaurant and its delivery service
- 00:22:41.3 - 00:22:51.7 Speaker 2 Parfait. Aussi, je voudrais que toutes les commandes soient suivies en temps réel. Euh, je veux savoir quand chaque plat quitte le restaurant et quand il arrive.
- 00:22:52.9 - 00:22:55.2 Speaker 1 L'application le permet déjà.
- 00:22:57.1 - 00:22:57.8 Speaker 1 Vous verrez
- 00:22:58.9 - 00:23:05.5 Speaker 1 l'état de chaque commande, l'itinéraire et l'heure estimée d'arrivée.
- 00:23:07.3 - 00:23:11.2 Speaker 2 Excellent. Et concernant le paiement pour cette grosse commande ?
- 00:23:13.3 - 00:23:20.2 Speaker 1 Comme d'habitude, la commission est prélevée automatiquement et le reste vous sera
- 00:23:21.0 - 00:23:24.3 Speaker 1 versé vendredi par virement.
- 00:23:26.3 - 00:23:34.2 Speaker 2 Très bien, euh, mais j'aimerais que nous fassions un point a, euh, après cette journée pour analyser les problèmes éventuels.
Monologue: what artificial intelligence is for
- 00:01:51.1 - 00:01:53.7 Speaker 1 À quoi sert l'intelligence artificielle?
- 00:01:54.4 - 00:01:56.3 Speaker 1 L'intelligence artificielle permet
- 00:01:58.2 - 00:01:59.3 Speaker 1 d'analyser des textes,
- 00:02:01.6 - 00:02:02.4 Speaker 1 qu'ils soient
- 00:02:03.5 - 00:02:06.8 Speaker 1 oraux ou écrits. L'intelligence artificielle
- 00:02:07.6 - 00:02:13.9 Speaker 1 arrive de mieux en mieux à comprendre et utiliser le langage pour répondre automatiquement
- 00:02:14.8 - 00:02:15.7 Speaker 1 à des requêtes,
- 00:02:18.4 - 00:02:20.5 Speaker 1 à des requêtes variées. Aujourd'hui,
- 00:02:21.1 - 00:02:25.0 Speaker 1 elle est utilisée, par exemple, pour gérer les relations clients
- 00:02:25.6 - 00:02:25.7 Speaker 1 sur
- 00:02:26.2 - 00:02:27.8 Speaker 1 Internet ou par téléphone.
- 00:02:28.5 - 00:02:30.5 Speaker 1 Les agents conversationnels,
- 00:02:31.6 - 00:02:33.5 Speaker 1 ou chatbots en anglais,
- 00:02:34.5 - 00:02:35.2 Speaker 1 sont des systèmes
- 00:02:35.9 - 00:02:36.4 Speaker 1 intelligents
- 00:02:38.9 - 00:02:42.1 Speaker 1 qui arrivent à entretenir une conversation
- 00:02:42.8 - 00:02:44.0 Speaker 1 en langage naturel.
Want whole recordings? Ask for the free sample pack with complete conversations and monologues and their TXT and JSON transcripts. The same sample recordings, with their full transcripts, are on Hugging Face. The excerpts and transcripts are also on GitHub and Kaggle.
Sounds like what you need?
Talk it through with Andreas, our founder, in a 30-minute call. Tell him how many hours of French you need and what the model is for, and you will leave the call with a price. Download access follows within 24 hours of a signed order.
What is in the dataset
| Recording type | Recordings | Hours | Average |
|---|---|---|---|
| Conversation, business topicsTwo speakers role-play workplace situations such as client meetings, deliveries, sales and interviews | 919 | 505.9 | 33 min |
| Conversation, open topicsTwo speakers talk freely about daily life, football, music, education, religion and travel | 936 | 506.4 | 32 min |
| MonologueOne speaker talks on a topic of their choice, from technology to history | 221 | 109.4 | 30 min |
| Single speaker, type not labelledMono recordings uploaded without a recording type | 12 | 8.9 | 44 min |
| Total | 2,088 | 1,130.6 |
Recordings from France make up 27 hours. If you need speakers from France, Belgium, Switzerland or Canada, we can record them to order.
Where the speakers are
| Country | Hours |
|---|---|
| Algeria | 487 |
| Madagascar | 202 |
| Morocco | 139 |
| Rwanda | 105 |
| Benin | 95 |
| France | 27 |
| Côte d'Ivoire | 25 |
| Senegal | 13 |
| Burundi | 12 |
| Other or not stated | 26 |
Speakers' home regions include Antananarivo, Algiers, Béjaïa, Batna and Casablanca.
What it is good for
- Training and testing French ASR (speech recognition) on unscripted speech, not read sentences.
- Speaker diarization: each speaker in a conversation has their own channel, so turns and overlaps can be read from the channels.
- Voice agents and turn-taking: real interruptions, backchannels and topic changes between two people.
Specifications
| Language | French |
|---|---|
| In stock | 1,131 hours, counted 18 November 2026 |
| Speakers | 82 registered contributors, 27 of whom recorded monologues. Their conversation partners were not registered, so there are more than 82 distinct voices; that number is not counted. |
| Recorded | 24 January 2026 to 26 February 2026 |
| Audio | 48 kHz WAV. Conversations: stereo, one speaker per channel, plus one file per speaker. Single-speaker recordings: mono. Bit depth is 16 or 24 depending on the recording. |
| Devices and setting | 96% recorded on a laptop or PC and about 4% on a phone, in quiet indoor settings. |
| Transcripts | Machine transcripts with speaker labels, segment start and end times and word-level timestamps, as TXT and JSON, for 1,129 of the 1,131 hours. Not human-reviewed. |
| Topics | More than 1,700 distinct topic labels, from business meetings, interviews and sales to football, music, religion and daily life. |
| Licence | Commercial use under a written licence. Voice cloning is excluded. What to check in a speech data licence |
| Delivery | Download access within 24 hours of a signed order |
| Price | $35 per audio hour for all 1,131 hours ($39,585); a part of it is priced on a call |
Buying for a model? Book a call with the founder to agree the hours, the licence and the price for the full set or the part you need. Our guide to buying AI training data lists what to check in a sample before you sign.
Get the full French sample pack
Complete sample conversations and monologues with their TXT and JSON transcripts, sent to your inbox. Already know how many hours you need? Book a call with the founder and agree the price on the call.
Frequently asked questions
Is the French speech dataset available now?
Yes. 1,131 hours are in stock, counted in our recording database on 18 November 2026. You get download access within 24 hours of a signed order.
What does the French speech dataset contain?
1,855 two-person conversations (1,012 hours) and 233 single-speaker recordings (118 hours), recorded between 24 January 2026 and 26 February 2026. More than 1,700 distinct topic labels, from business meetings, interviews and sales to football, music, religion and daily life.
Who are the speakers?
82 registered contributors, 27 of whom recorded monologues. Their conversation partners were not registered, so there are more than 82 distinct voices; that number is not counted. Speakers' home regions include Antananarivo, Algiers, Béjaïa, Batna and Casablanca.
What transcripts come with it?
Machine transcripts with speaker labels, segment start and end times and word-level timestamps, as TXT and JSON, for 1,129 of the 1,131 hours. Not human-reviewed. If you need human-verified transcripts, we can quote a review pass.
What audio format is it delivered in?
48 kHz WAV. Conversations: stereo, one speaker per channel, plus one file per speaker. Single-speaker recordings: mono. Bit depth is 16 or 24 depending on the recording.
Can I use the French recordings for commercial model training?
Yes, under a written commercial licence that names the recordings you buy and the permitted uses. Speech recognition, evaluation and speech understanding are the usual uses. Voice cloning is excluded.
What does the French speech dataset cost?
$35 per audio hour when you license all 1,131 hours, so $39,585 for the full dataset. For part of it, such as a fixed number of hours or one recording type, book a 30-minute call with the founder and agree the price on the call.
How is it different from free French corpora?
Common Voice and FLEURS are short sentences read aloud, and MLS is audiobooks read aloud. These recordings are unscripted: people talking to each other or one person talking freely for 32 minutes on average, which is closer to what assistants, meeting tools and call analytics have to transcribe.
What should I check when evaluating French speech data?
Conversational French may omit "ne" in a negative phrase and shorten "il y a" to "y'a". Liaison and silent endings also make the link between a recording and its spelling less direct. Inspect the sample and reference transcript for the register, region and transcription conventions your evaluation uses.