lijiaoqiao/llm-gateway-competitors/litellm-wheel-src/litellm/llms/openai/transcriptions/guardrail_translation

OpenAI Audio Transcription Guardrail Translation Handler

Handler for processing OpenAI's audio transcription endpoint (/v1/audio/transcriptions) with guardrails.

Overview

This handler processes audio transcription responses by:

1. Applying guardrails to the transcribed text output
2. Returning the input unchanged (since input is an audio file, not text)

Data Format

Input Format

The input is an audio file, which cannot be guardrailed (it's binary data, not text).

{
  "model": "whisper-1",
  "file": "<audio file>",
  "response_format": "json",
  "language": "en"
}

Output Format

{
  "text": "This is the transcribed text from the audio file."
}

Or with additional metadata:

{
  "text": "This is the transcribed text from the audio file.",
  "duration": 3.5,
  "language": "en"
}

Usage

The handler is automatically discovered and applied when guardrails are used with the audio transcription endpoint.

Example: Using Guardrails with Audio Transcription

curl -X POST 'http://localhost:4000/v1/audio/transcriptions' \
-H 'Authorization: Bearer your-api-key' \
-F 'file=@audio.mp3' \
-F 'model=whisper-1' \
-F 'guardrails=["pii_mask"]'

The guardrail will be applied to the output transcribed text only.

Example: PII Masking in Transcribed Text

curl -X POST 'http://localhost:4000/v1/audio/transcriptions' \
-H 'Authorization: Bearer your-api-key' \
-F 'file=@meeting_recording.mp3' \
-F 'model=whisper-1' \
-F 'guardrails=["mask_pii"]' \
-F 'response_format=json'

If the audio contains: "My name is John Doe and my email is john@example.com"

The transcription output will be: "My name is [NAME_REDACTED] and my email is [EMAIL_REDACTED]"

Example: Content Moderation on Transcriptions

curl -X POST 'http://localhost:4000/v1/audio/transcriptions' \
-H 'Authorization: Bearer your-api-key' \
-F 'file=@audio.wav' \
-F 'model=whisper-1' \
-F 'guardrails=["content_moderation"]'

Implementation Details

Input Processing

• Status: Not applicable
• Reason: Input is an audio file (binary data), not text
• Result: Request data returned unchanged

Output Processing

• Field: text (string)
• Processing: Applies guardrail to the transcribed text
• Result: Updated text in response

Use Cases

1. PII Protection: Automatically redact personally identifiable information from transcriptions
2. Content Filtering: Remove or flag inappropriate content in transcribed audio
3. Compliance: Ensure transcriptions meet regulatory requirements
4. Data Sanitization: Clean up transcriptions before storage or further processing

Extension

Override these methods to customize behavior:

• process_output_response(): Customize how transcribed text is processed
• process_input_messages(): Currently a no-op, but can be overridden if needed

Supported Call Types

• CallTypes.transcription - Synchronous audio transcription
• CallTypes.atranscription - Asynchronous audio transcription

Notes

• Input processing is a no-op since audio files cannot be text-guardrailed
• Only the transcribed text output is processed
• Guardrails apply after transcription is complete
• Both sync and async call types use the same handler
• Works with all Whisper models and response formats

Common Patterns

Transcribe and Redact PII

import litellm

response = litellm.transcription(
    model="whisper-1",
    file=open("interview.mp3", "rb"),
    guardrails=["mask_pii"],
)

# response.text will have PII redacted
print(response.text)

Async Transcription with Guardrails

import litellm
import asyncio

async def transcribe_with_guardrails():
    response = await litellm.atranscription(
        model="whisper-1",
        file=open("audio.mp3", "rb"),
        guardrails=["content_filter"],
    )
    return response.text

text = asyncio.run(transcribe_with_guardrails())