Dify audio transcription and synthesis

Use this family to convert one authorized audio file to text or convert
authorized text to speech. Both operations use the selected app credential and
a stable Dify end-user identity.

Audio content can contain sensitive personal data. Approval must cover the
source material, selected app, end user, model invocation, and intended result
handling before either request runs.

Choose an operation

| Operation suffix | Method and route | Request | Result | Connector projection |

| audio.transcribe | POST /v1/audio-to-text | Multipart file and user | JSON transcript | All six app modes |
| audio.synthesize | POST /v1/text-to-audio | JSON body and user | Bounded binary | All six app modes |

Both operations are medium-risk Tool mutations. Each requires tenant-policy
approval and one Action idempotency key, publishes no safe retry or
cancellation, and declares rollback unsupported.

The capability template is:

cap:dify::

Use tenant discovery to select the app capability. The Action cannot replace
the configured provider origin or Bearer key.

Transcribe one audio file

Before dispatch, verify that the audio belongs to the selected end user and
that the chosen Dify app enables speech-to-text. Supply a safe filename, exact
MIME type, and base64 content:

{
  "user": "tenant-user-0190",
  "file": {
    "filename": "request.wav",
    "contenttype": "audio/wav",
    "contentbase64": "UklGRg=="
  }
}

The connector decodes the content and sends multipart fields named file and
user. It does not send Action body as a multipart field for this operation.

The pinned provider accepts these MIME types:
• audio/mp3
• audio/mpga
• audio/m4a
• audio/wav
• audio/amr

The provider limit is 30 MiB. The connector permits at most 32 MiB of decoded
upload data, while its base64 schema and the native Action envelope can impose
a lower effective ceiling. The smallest applicable limit governs the request.

A successful result uses the generic JSON envelope:

{
  "httpstatus": 200,
  "body": {
    "text": "The transcribed speech appears here."
  },
  "contenttype": "application/json",
  "providerrequestid": null
}

The connector does not normalize the language, confidence, speaker identity,
or punctuation. At the pinned provider revision, the public response contains
one required text value.

Synthesize bounded speech

The provider accepts messageid, voice, text, and streaming in the
operation body. The connector inserts the validated user value.

Use direct text when the caller owns the content:

{
  "user": "tenant-user-0190",
  "body": {
    "text": "Your scheduled review begins at fourteen thirty.",
    "voice": "provider-voice-id"
  }
}

When messageid is present, Dify gives that message priority over text.
The message lookup is constrained to the selected app and end user. Without a
message ID, text is required.

voice is optional. Dify uses the app's configured voice when available, or a
voice returned by its model provider. The connector does not publish a voice
catalog and does not validate provider-specific voice identifiers.

The streaming field is reserved by Dify for compatibility. Provider output
can arrive as streamed audio/mpeg, but this AIP operation buffers the bytes
and returns one bounded binary result.

{
  "httpstatus": 200,
  "contentbase64": "SUQz",
  "contenttype": "audio/mpeg",
  "sizebytes": 3,
  "providerrequestid": null
}

The configured connector response limit applies to all returned audio. Its
default is 8 MiB, with an allowed configuration range from 1 to 64 MiB.

Decode the result only after authorization, content-type policy, malware
controls, and output-size checks. The connector does not write audio to a
caller-selected filesystem path or preserve it as durable provider state.

Apply mutation governance

Each operation requires a non-blank Action idempotency key of at most 512
bytes. AIP retains the tenant-scoped reservation for 24 hours and revalidates
the input hash on collision.

The key does not prove that Dify or the selected model provider deduplicates
the invocation. Neither operation publishes automatic retry, transaction,
reconciliation, cancellation, or compensation support.

Treat the uploaded audio, source text, referenced message, transcript, and
generated bytes as confidential PII-bearing data. Retain only the evidence
needed to establish authorization and outcome.

Recover safely

| Observation | Decision |

| Filename, base64, user, or body is rejected before dispatch | Correct the same request without claiming a provider effect |
| Provider rejects the audio MIME type | Convert or select an allowed source under a new reviewed input |
| Audio exceeds an effective request limit | Reduce the authorized source; do not raise response limits |
| Speech-to-text or text-to-speech is disabled | Correct the selected app configuration before a new Action |
| Message lookup returns no usable audio | Verify the same app, end user, message ID, and message state |
| Voice is absent or unsupported | Select a provider-advertised voice or repair app configuration |
| Binary output exceeds the response bound | Treat the result as failed and review expected output size |
| Response is lost after model invocation | Treat the external outcome as uncertain; do not replay automatically |
| Returned transcript or audio crosses an app or user boundary | Stop traffic and treat the event as an isolation incident |

Preserve the Action ID, input digest, approval, idempotency key, provider
request identity, and any bounded result used to settle an ambiguous outcome.

Related documentation
• Dify capability index (README.md)
• Dify app metadata and files (app-metadata-and-files.md)
• Dify connector configuration (../reference/configuration.md)
• Authentication and credential scopes (../getting-started/authentication-and-credential-scopes.md)
• Errors and retry decisions (../../../reference/errors.md)