> ## Documentation Index
> Fetch the complete documentation index at: https://dev.docs.inworld.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Zero Data Retention

> Enterprise-grade data privacy across Inworld TTS, STT, and LLMs with SOC 2 Type 2, HIPAA, and GDPR compliance

Enterprise customers in healthcare, finance, and other regulated industries can ensure that eligible customer inputs and generated outputs are not persistently stored in our systems after processing completes. Zero Data Retention is available across Inworld TTS, STT, and hosted LLMs. When enabled, eligible requests are processed in memory and their inputs and outputs are immediately redacted from our logging systems once processing completes. Zero Data Retention is configured at the workspace level. What is covered, and the exceptions that remain, varies by product, see the sections below for details.

## TTS

When enabled, all customer text inputs sent to our text-to-speech engine are processed in memory to generate audio. Once complete, both the text and the generated audio are immediately redacted from our logging systems. This applies to all text sent for speech synthesis, including requests using cloned voices.

| Product        | Type                     | Eligible for Zero Data Retention? |
| :------------- | :----------------------- | :-------------------------------- |
| Text-to-Speech | Text Input               | Yes                               |
| Text-to-Speech | Audio Output             | Yes                               |
| Voice Cloning  | Audio Samples            | No                                |
| Voice Design   | Voice Description/Script | No                                |

## STT

When a request is processed, customer audio is sent to the selected provider and transcribed in memory. Once complete, both the original audio and the transcription output are immediately redacted from our logging systems. This applies to requests, both sync (file upload) and WebSocket (streaming), processed through the Inworld STT API. Each upstream provider maintains its own guarantees, which can be mode-specific, so see the per-provider details below.

| Data                 | Type                                      | Eligible for Zero Data Retention? |
| :------------------- | :---------------------------------------- | :-------------------------------- |
| Audio Input          | Customer-submitted audio (file or stream) | Yes                               |
| Transcription Output | Transcribed text returned by the API      | Yes                               |
| Usage Metadata       | Request duration, model ID, timestamps    | No (retained for billing)         |

Zero Data Retention applies to all STT models available through the Inworld STT API. Each upstream provider maintains its own ZDR guarantees:

| **Provider** | **Models**                                                | **Details**                                                                                                                       |
| :----------- | :-------------------------------------------------------- | :-------------------------------------------------------------------------------------------------------------------------------- |
| Inworld      | `inworld/inworld-stt-1`                                   | Processed in memory; redacted from logs after completion                                                                          |
| AssemblyAI   | `assemblyai/*`                                            | ZDR enabled for streaming ([AssemblyAI docs](https://www.assemblyai.com/docs/data-retention-and-model-training))                  |
| Groq         | `groq/whisper-large-v3`                                   | No customer inputs or outputs retained by default ([Groq docs](https://console.groq.com/docs/your-data))                          |
| Soniox       | `soniox/stt-rt-v4`, `soniox/stt-rt-v5`                    | Processed in memory; redacted from logs after completion ([Soniox docs](https://soniox.com/docs/stt/api-reference/websocket-api)) |
| Deepgram     | `deepgram/flux-general-en`, `deepgram/flux-general-multi` | No customer audio retained by default ([Deepgram docs](https://developers.deepgram.com/docs/data-privacy-compliance))             |

## LLM Router

<Note>
  Zero Data Retention is currently available only for Inworld hosted LLM models. Requests routed to third-party model providers are not covered.
</Note>

When enabled, all customer prompts sent to Inworld hosted LLMs are processed in memory to generate a response. Once complete, neither the prompt nor the generated output is written to our logging or storage systems. This applies to all content sent for completion, including system prompts and tool interactions.

| Product                     | Type                      | Eligible for Zero Data Retention? |
| :-------------------------- | :------------------------ | :-------------------------------- |
| Inworld hosted LLM          | Prompt / Input messages   | Yes                               |
| Inworld hosted LLM          | Model output / Completion | Yes                               |
| Third-party model providers | Prompt & Output           | No                                |

## Enterprise-level security

Inworld's platform is SOC 2 Type II certified and HIPAA and GDPR compliant. For full details on our security posture, certifications, and data handling practices, see our [Security page](https://inworld.ai/security).

<CardGroup cols={2}>
  <Card title="Workspace Configuration" icon="shield-check">
    Zero Data Retention is configured at the workspace level, allowing customers with strict data privacy requirements to maintain compliance.
  </Card>

  <Card title="Production Scale" icon="gear">
    Whether you're handling protected health information or sensitive financial data, you get the security guarantees needed for production deployment at scale.
  </Card>
</CardGroup>

## FAQs

<AccordionGroup>
  <Accordion title="Can we enable Zero Data Retention for only specific workspaces?">
    Yes. Zero Data Retention is configured per workspace. Contact our sales team to enable it for specific workspaces.
  </Accordion>

  <Accordion title="How does Zero Data Retention affect debugging?">
    With Zero Data Retention enabled, original inputs and outputs cannot be retrieved. Debugging and troubleshooting capabilities are limited to non-sensitive metadata only.
  </Accordion>

  <Accordion title="How does Inworld TTS Zero Data Retention work?">
    TTS Zero Data Retention protects privacy by redacting all input texts from logs, including text sent to Cloned Voices, while only retaining the initial source audio required to build the voice itself. This is currently a workspace-level configuration managed by the Inworld Engineering team.
  </Accordion>

  <Accordion title="Does TTS Zero Data Retention affect audio output quality?">
    No. The TTS engine still receives the text to generate the audio. The redaction happens only at the logging layer, ensuring the record of what was said is not saved.
  </Accordion>

  <Accordion title="How does Inworld STT Zero Data Retention work?">
    STT Zero Data Retention protects privacy by redacting all customer audio and transcription outputs from logs immediately after processing. Only non-sensitive usage metadata (request duration, model ID, timestamps) is retained for billing and service monitoring.
  </Accordion>

  <Accordion title="Does Zero Data Retention apply to all STT models?">
    Yes. All models available through the Inworld STT API operate under Zero Data Retention. Each upstream provider's specific guarantees are listed in the supported models table above.
  </Accordion>

  <Accordion title="Is my audio used for model training?">
    No. Audio processed through the Inworld STT API is never used for training by Inworld or any upstream provider.
  </Accordion>

  <Accordion title="Does STT Zero Data Retention affect transcription quality?">
    No. The STT engine still receives the audio to generate the transcript. The redaction happens only at the logging layer, ensuring no record of the audio or transcript is saved.
  </Accordion>

  <Accordion title="How does Inworld LLM Zero Data Retention work?">
    LLM Zero Data Retention protects privacy by never writing prompts or model outputs to our logs. It is currently available for Inworld hosted LLM models, and is a workspace-level configuration managed by the Inworld Engineering team.
  </Accordion>

  <Accordion title="Which LLM models support Zero Data Retention?">
    Zero Data Retention is currently available only for Inworld hosted LLM models. Requests routed to third-party model providers are not covered.
  </Accordion>

  <Accordion title="Does LLM Zero Data Retention affect output quality?">
    No. The model still receives the full prompt to generate its response. Zero Data Retention applies only at the logging layer: the prompt and output are never written there, so no record of the conversation is saved.
  </Accordion>
</AccordionGroup>