Prestataires de services

Voici les sociétés qui traitent des données personnelles pour Grasp, ce que chacune d'elles reçoit et les engagements qui s'appliquent. Notre Politique de confidentialité explique les types de prestataires auxquels nous faisons appel et vos droits.

Plusieurs prestataires peuvent assurer une même tâche, par exemple un prestataire principal et un prestataire de secours, ou différents prestataires selon les langues ou les pays, et nos systèmes basculent automatiquement de l'un à l'autre. Chacun d'eux est indiqué ici et soumis aux mêmes règles.

1. Hébergement

Cloudflare

• Société: Cloudflare, Inc. (USA)
• Ce qu'il fait: Hosting of the site and the API, database, file storage including our encrypted backups, network protection, bot checks (Turnstile), visit statistics on our public pages (Web Analytics), and the relay for direct calls between devices (Cloudflare Realtime TURN)
• Données: The data Grasp stores, our encrypted backups, IP addresses, addresses and technical details of visits to public pages; for direct calls it relays, the encrypted call audio in transit and the two devices' IP addresses; for Talk in person, the text it passes between phones and, in the ten new call languages, the microphone audio it passes to their recognizers, in transit only
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Keeps our data only for us and for as long as our retention periods say (Privacy Policy, section 8)
• Utilisation: Main; all countries

Google Cloud

• Société: Google LLC (USA)
• Ce qu'il fait: Hosts the servers in the United States that run our speech-processing service, including our own noise filter; we operate these servers ourselves
• Données: The audio and text our speech-processing service processes (call audio and captions, including calls to phone numbers and their notices, voice messages, the sound of video messages and their transcripts, voice recordings, the voice of Talk in person, display names and gender form used as hints), held in memory only; server logs with internal IDs
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Audio and text are held in memory only and never written to disk; server logs are deleted within 30 days
• Utilisation: Main; all countries

2. Transport des appels

LiveKit

• Société: LiveKit, Inc. (USA)
• Ce qu'il fait: Real-time audio and video transport for calls; through LiveKit Inference, passes the translated text of a speaker with a standard voice to Inworld's ready-made voices, and the audio of voice messages and the sound of video messages to Deepgram for speech recognition; for calls to phone numbers, its telephone service (LiveKit SIP) connects the call to Amazon Web Services; through LiveKit Inference, passes speech in Icelandic, Nepali and Mongolian to OpenAI and Speechmatics for speech recognition
• Données: Call audio, video and captions in transit, display name, internal participant ID, IP address; for a standard voice, the translated text and its language; the audio of voice messages and the sound of video messages sent to Deepgram; for a call to a phone number, the number called, the Grasp number the person sees and the audio of the phone line; speech in Icelandic, Nepali and Mongolian, in calls and in voice and video messages
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Under LiveKit's terms, what passes through LiveKit Inference is not kept beyond each request and not used to train models. LiveKit keeps records of calls to phone numbers, which hold the number called, as its terms set: as long as reasonably necessary and as the law or phone carriers' rules require; it publishes no fixed period
• Utilisation: Main for calls; all countries

3. Réseau téléphonique

Amazon Web Services (Amazon Chime SDK Voice Connector)

• Société: Amazon Web Services, Inc., with its affiliate AMCS LLC, which sells the service (USA)
• Ce qu'il fait: Carries calls to phone numbers from LiveKit to the phone network and completes them through phone carriers; gives the Grasp phone number the person called sees
• Données: The number called, the Grasp number the person sees, and the audio of the phone line: the translations the person hears and what they say
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Acts as our processor under the AWS data processing terms. We have not switched on its call records or its media streaming, so it keeps neither for us; its documents do not say how long it keeps its own records of calls
• Utilisation: Main for calls to phone numbers, made by people outside the United States, the EU, the EEA, the United Kingdom and Switzerland

Phone carriers

• Société: The carriers Amazon Web Services works with, and the phone company of the number called (The country of the number called, and the countries on the way)
• Ce qu'il fait: Complete calls to phone numbers in the phone network of the number called
• Données: The number called, the Grasp number the person sees, the audio of the phone line and the times of the call
• Rôle: Independent controller
• Garantie: Necessary to complete a call the caller asked for (GDPR Article 49(1)(b)); their own safeguards and the law of their country
• Conservation et entraînement: Keep their own records of calls under their own rules and the law of their country
• Utilisation: Every call to a phone number, as the phone network routes it

4. Reconnaissance et traduction de la parole

Soniox

• Société: Soniox, Inc. (USA)
• Ce qu'il fait: Speech recognition and translation in calls, including calls to phone numbers, and in Talk in person; speech in standard voices; the clones of voices it cloned (voices created before 29 October 2026); speech recognition of voice messages and of the sound of video messages when neither AssemblyAI nor Deepgram is used for them
• Données: Call audio, Talk in person audio and, when you use Talk in person, your IP address; voice messages and the sound of video messages in those languages; the recordings and clones of the voices it cloned; display names and your gender form, as hints
• Rôle: Processor
• Garantie: Standard Contractual Clauses (controller to processor) with the UK Addendum, in its data processing agreement
• Conservation et entraînement: States in its terms that it does not use its customers' content to train or improve its models
• Utilisation: Main for calls and Talk in person; all countries

AssemblyAI

• Société: AssemblyAI, Inc. (USA)
• Ce qu'il fait: Speech recognition of voice messages and of the sound of video messages; not used in calls
• Données: The audio of voice messages and the sound of video messages, and the language of the speech when we know it
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Our account is opted out of its model improvement; it deletes the audio and the transcript within 1 day
• Utilisation: Main for voice and video messages, in the languages it supports; all countries

Deepgram

• Société: Deepgram, Inc. (USA)
• Ce qu'il fait: Recognizes speech in voice messages and in the sound of video messages, in languages AssemblyAI does not support
• Données: Those voice messages and the sound of those video messages, and the language of the speech
• Rôle: Sub-processor, through LiveKit
• Garantie: LiveKit's terms with its providers
• Conservation et entraînement: Under LiveKit's terms, nothing is stored and nothing is used to train models
• Utilisation: Second for voice and video messages, in the languages it supports; not for people in Russia, Belarus, mainland China or Syria

Inworld (speech recognition)

• Société: Theai, Inc. (Inworld AI) (USA)
• Ce qu'il fait: Recognizes Cantonese speech in calls, in Talk in person, in voice messages and in the sound of video messages
• Données: Call audio, Talk in person audio, the audio of voice messages and the sound of video messages in Cantonese
• Rôle: Processor
• Garantie: Standard Contractual Clauses in its data processing addendum
• Conservation et entraînement: Acts as our processor under its data processing addendum; its documents do not say how long it keeps the audio and text of speech recognition
• Utilisation: Cantonese only, once that language is switched on; all countries

OpenAI (speech recognition model)

• Société: OpenAI, L.L.C. (USA)
• Ce qu'il fait: Recognizes Icelandic, Nepali and Armenian speech in calls, in Talk in person, in voice messages and in the sound of video messages
• Données: Call audio, Talk in person audio, the audio of voice messages and the sound of video messages in those languages
• Rôle: Sub-processor, through LiveKit
• Garantie: LiveKit's terms with its providers
• Conservation et entraînement: Under LiveKit's terms, nothing is kept beyond each request and nothing is used to train models
• Utilisation: Icelandic, Nepali and Armenian only, once those languages are switched on; all countries

Speechmatics (speech recognition model)

• Société: Cantab Research Limited, trading as Speechmatics (United Kingdom; LiveKit may have the speech recognized in the United States, the United Kingdom or the EU)
• Ce qu'il fait: Recognizes Mongolian speech in calls, in Talk in person, in voice messages and in the sound of video messages
• Données: Call audio, Talk in person audio, the audio of voice messages and the sound of video messages in Mongolian
• Rôle: Sub-processor, through LiveKit
• Garantie: LiveKit's terms with its providers, which cover every provider reached through LiveKit Inference; the United Kingdom has an EU adequacy decision
• Conservation et entraînement: Under LiveKit's terms, nothing is kept beyond each request and nothing is used to train models
• Utilisation: Mongolian only, once that language is switched on; all countries

Amazon Web Services (Amazon Transcribe)

• Société: Amazon Web Services, Inc. (USA)
• Ce qu'il fait: Recognizes Georgian, Odia, Uzbek, Burmese and Javanese speech in calls, in Talk in person, in voice messages and in the sound of video messages
• Données: Call audio, Talk in person audio, the audio of voice messages and the sound of video messages in those languages
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Acts as our processor under the AWS data processing terms. We have set the AWS AI services opt-out policy for our organization, so under AWS's documentation it does not store or use this audio or its text to develop or improve its services
• Utilisation: For the new call languages Georgian, Odia, Uzbek, Burmese and Javanese, once each is switched on; all countries

5. Synthèse vocale et clonage de voix

Inworld

• Société: Theai, Inc. (Inworld AI) (USA)
• Ce qu'il fait: Makes the clone of your voice and speaks your translated speech and the previews of your voice in it; its ready-made voices speak for people with a standard voice, directly or reached through LiveKit, and the notice of a call to a phone number
• Données: A part of your voice recording and the clone made from it; in calls, including calls to phone numbers, and with voice on in Talk in person, the translated text of what you say and its language; the text of the notice of a call to a phone number, with the caller's display name; the sentences of the previews of your voice
• Rôle: Processor
• Garantie: Standard Contractual Clauses in its data processing addendum
• Conservation et entraînement: Acts as our processor under its data processing addendum; its documents do not say how long it keeps call text in its logs
• Utilisation: Main for personal voices and for standard voices in calls; every language but Bosnian; all countries

6. Modèles de langage d'IA, directement ou via un routeur de modèles

OpenRouter

• Société: OpenRouter, Inc. (USA)
• Ce qu'il fait: Model router: sends our text to the AI models below
• Données: Chat messages to translate with a few earlier messages of the same chat, voice and video message transcripts, text typed in Talk in person, text of voice tests, call phrases to rewrite as statements, and the recognized speech of calls in Cantonese, Icelandic, Nepali, Mongolian, Georgian, Odia, Uzbek, Burmese, Javanese and Armenian
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Every request is sent with data collection denied and zero data retention requested
• Utilisation: Backup for chat translation, text typed in Talk in person and calls in Cantonese, Icelandic, Nepali, Mongolian, Georgian, Odia, Uzbek, Burmese, Javanese and Armenian; main for voice tests and rewriting call phrases; all countries

Microsoft Azure (OpenAI model)

• Société: Microsoft Corporation (USA)
• Ce qu'il fait: Hosts the OpenAI model that translates text, and the backup for rewriting call phrases
• Données: The text OpenRouter sends
• Rôle: Sub-processor, through OpenRouter
• Garantie: OpenRouter's terms with its providers
• Conservation et entraînement: Zero data retention requested with every request
• Utilisation: Backup for chat translation and text typed in Talk in person; main for voice tests; outside Russia, Belarus, mainland China and Syria

Google (Gemini model)

• Société: Google LLC (USA)
• Ce qu'il fait: Hosts the Gemini model: translates text when the main host cannot answer, and rewrites call phrases first
• Données: The text OpenRouter sends
• Rôle: Sub-processor, through OpenRouter
• Garantie: OpenRouter's terms with its providers
• Conservation et entraînement: Zero data retention requested with every request
• Utilisation: Backup for text translation, main for rewriting call phrases; outside Russia, Belarus, mainland China and Syria

DeepInfra, Together AI, Baseten, Fireworks AI (DeepSeek model)

• Société: DeepInfra, Together AI, Baseten, Fireworks AI (USA)
• Ce qu'il fait: Host an open model (DeepSeek) where OpenAI and Google models are not available, tried in this order
• Données: The text OpenRouter sends
• Rôle: Sub-processor, through OpenRouter
• Garantie: OpenRouter's terms with its providers
• Conservation et entraînement: Zero data retention requested with every request
• Utilisation: Backup for chat translation and text typed in Talk in person; main for voice tests; for people located in Russia, Belarus, mainland China or Syria, or with one of them as profile country

Amazon Web Services (Google Gemma model)

• Société: Amazon Web Services, Inc. (USA)
• Ce qu'il fait: Amazon Bedrock runs Google's open model Gemma, which translates chat messages, text typed in Talk in person and the speech of calls and of Talk in person in Cantonese, Icelandic, Nepali, Mongolian, Georgian, Odia, Uzbek, Burmese, Javanese and Armenian
• Données: Chat messages to translate with a few earlier messages of the same chat, voice and video message transcripts to translate, text typed in Talk in person, and in calls in those languages the recognized text of what each person says, with up to two earlier phrases of the same speaker
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Under AWS's documentation, Amazon Bedrock keeps prompts and answers only for the models it names for abuse detection, which do not include Gemma, and does not share them with model providers
• Utilisation: Main for chat translation, text typed in Talk in person and calls in Cantonese, Icelandic, Nepali, Mongolian, Georgian, Odia, Uzbek, Burmese, Javanese and Armenian; all countries

Amazon Web Services (Amazon Translate)

• Société: Amazon Web Services, Inc. (USA)
• Ce qu'il fait: Translates the speech of calls and of Talk in person into Armenian
• Données: In calls and Talk in person into Armenian, the recognized text of what each person says
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Acts as our processor under the AWS data processing terms. We have set the AWS AI services opt-out policy for our organization, so under AWS's documentation it does not store or use this text or its translation to develop or improve its services
• Utilisation: Main for calls into Armenian, once Armenian is switched on; all countries

7. E-mail

Cloudflare Email

• Société: Cloudflare, Inc. (USA)
• Ce qu'il fait: Sends our emails, such as sign-in codes, invitations and notices
• Données: The email address and message of each email we send
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Delivers our emails for us only
• Utilisation: Main; all countries

8. Notifications push

Apple Push Notification service, Firebase Cloud Messaging

• Société: Apple Inc., Google LLC (USA)
• Ce qu'il fait: Push notifications in our mobile apps, once they are available
• Données: Device token, the caller's display name for call alerts
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Deliver notifications only
• Utilisation: Main for our mobile apps, once they are available

Your browser's push service

• Société: For example Google LLC, Mozilla Corporation or Apple Inc. (USA)
• Ce qu'il fait: Delivers web push notifications to the browser you use
• Données: The push endpoint and the notification, which we encrypt for your browser, so the service cannot read it
• Rôle: Processor
• Garantie: We encrypt the content for your browser before it leaves our servers, so the service cannot read it
• Conservation et entraînement: Carries content it cannot read
• Utilisation: Chosen by your browser

9. Assistance

Google Workspace

• Société: Google LLC (USA)
• Ce qu'il fait: Our company email, including the mailbox support@grasp.fm
• Données: What you write to us; your email address and our message when we reply or send an invitation from our mailbox
• Rôle: Processor
• Garantie: Data Privacy Framework where the provider is certified, otherwise Standard Contractual Clauses in its data processing terms
• Conservation et entraînement: Keeps our mail for us; we keep a conversation up to 3 years after its last message (Privacy Policy, section 8)
• Utilisation: Main; all countries

10. Paiements

Stripe and Link

• Société: Stripe, Inc. and its affiliates (USA)
• Ce qu'il fait: Sells Grasp Plus, Grasp Lite and top-ups of the phone call balance on the web as the merchant of record, and handles payment, tax and refunds
• Données: Your email address and an internal account ID when you start a purchase; your payment details, which we never receive
• Rôle: Independent controller
• Garantie: Stripe's own safeguards, as a controller under its own privacy policy
• Conservation et entraînement: Under its own privacy policy
• Utilisation: Main for web purchases, where paid plans are offered

11. Connexion

Sign in with Apple, Google sign-in

• Société: Apple Inc., Google LLC (USA)
• Ce qu'il fait: Sign you in with your Apple or Google account, if you choose it
• Données: Your name, email address (or Apple's private relay address) and a stable account identifier
• Rôle: Independent controller
• Garantie: Their own safeguards, as controllers under their own privacy policies
• Conservation et entraînement: Under their own privacy policies
• Utilisation: Only if you choose it

12. Journal des modifications

• 4 November 2026: Talk in person speaks the ten new call languages too, once each is switched on for an account: Cloudflare passes the microphone audio of a speaker of one of them, in transit only, to the recognizer calls use for that language (Amazon Transcribe; OpenAI or Speechmatics through LiveKit Inference; Inworld for Cantonese), and Gemma on Amazon Bedrock, or Amazon Translate into Armenian, translates what was said.
• 4 November 2026: Inworld's ready-made voices speak the notice of a call to a phone number, with the caller's display name, instead of Soniox's; for a Bosnian number a Croatian or Serbian voice of Inworld reads the Bosnian text.
• 3 November 2026: Inworld speaks every translated voice in calls: a person's own clone keeps its place for the whole call, and Inworld's ready-made voices speak for people with a standard voice directly too, not only through LiveKit. Soniox's voices speak in calls only for an older personal voice made at Soniox and for Bosnian, which Inworld does not speak.
• 3 November 2026: Armenian becomes a call language, once it is switched on: OpenAI, through LiveKit Inference, recognizes it, as it does Icelandic and Nepali; Amazon Web Services translates the speech of calls into Armenian with Amazon Translate in the USA, as our processor, under the same AWS AI services opt-out; Gemma on Amazon Bedrock translates Armenian speech into the other languages.
• 3 November 2026: For more new call languages, once each is switched on: Amazon Web Services recognizes Georgian, Odia, Uzbek, Burmese and Javanese speech with Amazon Transcribe in the USA, as our processor; we have set the AWS AI services opt-out policy, so AWS does not use this audio or its text to improve its services. Gemma on Amazon Bedrock translates the speech of calls in these languages, with OpenRouter as backup.
• 3 November 2026: A new kind, Telephone network, for calls to phone numbers: LiveKit connects such a call through its telephone service (LiveKit SIP) to Amazon Web Services (Amazon Chime SDK Voice Connector), which completes it through phone carriers in the country of the number called. The person called sees a Grasp number. LiveKit keeps records of these calls, which hold the number called, as its terms set. Soniox speaks the notice the person hears before answering. Stripe and Link also sell top-ups of the phone call balance.
• 3 November 2026: Talk in person replaces the earlier translator: the browser sends microphone audio directly to Soniox; typed text goes through the chat's text route, first to Gemma on Amazon Bedrock, with OpenRouter and the models it reaches as backup; with voice on, our speech-processing service speaks the translations through Inworld or Soniox. Cloudflare passes the session's text between the phones that joined it.
• 3 November 2026: For new call languages, once each is switched on: Inworld recognizes Cantonese; through LiveKit Inference, OpenAI recognizes Icelandic and Nepali, and Speechmatics, a UK company, recognizes Mongolian, which LiveKit may have recognized in the United States, the United Kingdom or the EU. Gemma on Amazon Bedrock translates the speech of calls in these languages, with OpenRouter as backup.
• 3 November 2026: The commitments of LiveKit Inference and of Amazon Bedrock now use the words of their current terms and documentation: LiveKit keeps nothing that passes through LiveKit Inference beyond each request; Amazon Bedrock keeps prompts and answers only for the models it names for abuse detection, which do not include Gemma.
• 31 October 2026: Calls without translation between contacts can now go directly between devices. Cloudflare relays such a call, through Cloudflare Realtime TURN, when the two devices cannot connect directly or when a person has switched direct connection off in Settings. It carries only the encrypted call audio and sees the two devices' IP addresses.
• 31 October 2026: Deepgram, through LiveKit, now recognizes the speech of voice messages and of the sound of video messages in the languages AssemblyAI does not support, except for people in Russia, Belarus, mainland China and Syria. It no longer serves calls.
• 31 October 2026: Chat messages and the text of voice and video messages are now translated by Google's open model Gemma, run by Amazon Web Services on Amazon Bedrock in the USA, for people in all countries. OpenRouter and the models it reaches are now the backup for chat translation and stay main for text typed in the translator, voice tests and rewriting call phrases.
• 31 October 2026: LiveKit also carries the video of calls when a person turns their camera on. Call video is not recorded or stored and never goes to our speech-processing service or any AI provider.
• 30 October 2026: This page published. The Privacy Policy and the Voice and AI notice now name kinds of providers and point here for each provider.