Skip to main content

​
Overview

Rubric accepts both audio recordings and transcripts for voice AI evaluation. This page covers supported formats, processing options, and best practices.

​
Audio Formats

​
Supported Formats

FormatExtensionNotes
WAV.wavRecommended for highest quality
MP3.mp3Good compression, widely supported
M4A.m4aAAC codec, Apple ecosystem
FLAC.flacLossless compression
OGG.oggOpen format
WebM.webmBrowser recordings

​
Quality Requirements

For best transcription accuracy, we recommend:
  • Sample rate: 16kHz or higher
  • Bit depth: 16-bit or higher
  • Channels: Mono or stereo with clear speaker separation

​
Providing Audio

# Option 1: Public URL
client.calls.log(
    audio_url="https://storage.example.com/calls/call_123.wav",
    ...
)

# Option 2: Signed URL (recommended for private storage)
client.calls.log(
    audio_url="https://bucket.s3.amazonaws.com/call.wav?X-Amz-Signature=...",
    ...
)

# Option 3: Upload directly
with open("call.wav", "rb") as f:
    client.calls.log(
        audio=f,
        ...
    )

​
Transcript Format

​
Standard Format

{
  "transcript": [
    {
      "speaker": "agent",
      "text": "Thank you for calling. How can I help you today?",
      "start": 0.0,
      "end": 2.5
    },
    {
      "speaker": "patient", 
      "text": "I've been having chest pain since this morning.",
      "start": 3.0,
      "end": 6.5
    }
  ]
}

​
Required Fields

​
speaker
string
required
Speaker identifier. Use agent for AI and patient for the caller. Custom labels supported.
​
text
string
required
The spoken text content.

​
Optional Fields

​
start
float
Start time in seconds from beginning of audio.
​
end
float
End time in seconds.
​
confidence
float
Transcription confidence score (0-1).
​
words
array
Word-level timestamps for fine-grained analysis.
{
  "words": [
    {"word": "chest", "start": 3.2, "end": 3.5, "confidence": 0.98},
    {"word": "pain", "start": 3.5, "end": 3.9, "confidence": 0.99}
  ]
}

​
Speaker Diarization

If you provide audio without a transcript, Rubric can perform automatic speaker diarization: 
client.calls.log(
    project="triage",
    audio_url="https://...",
    
    # Enable automatic transcription and diarization
    transcribe=True,
    diarize=True,
    
    # Hint: expected number of speakers
    expected_speakers=2,
    
    ai_decision={...}
)

​
Diarization Options

OptionDescription
expected_speakersHint for number of speakers (improves accuracy)
speaker_labelsMap speaker IDs to roles: {"SPEAKER_00": "agent", "SPEAKER_01": "patient"}

​
Transcript Sources

Rubric integrates with popular transcription services:

Deepgram

Real-time streaming transcription

AssemblyAI

High accuracy, medical vocabulary

Whisper

Open source, self-hosted option

​
Using Pre-transcribed Data

If you already have transcripts from your provider: 
# Deepgram format
deepgram_response = {...}  # From Deepgram API

client.calls.log(
    project="triage",
    transcript_source="deepgram",
    transcript_raw=deepgram_response,
    ai_decision={...}
)
Rubric will automatically normalize the format.

​
Audio-Transcript Alignment

When both audio and transcript are provided, Rubric can:
  1. Verify alignment - Check that transcript matches audio
  2. Identify gaps - Find segments missing from transcript
  3. Flag discrepancies - Detect potential transcription errors
client.calls.log(
    audio_url="https://...",
    transcript=[...],
    
    # Enable alignment verification
    verify_alignment=True,
    alignment_tolerance=0.5  # seconds
)

​
Privacy & Redaction

Voice recordings may contain PHI. Ensure proper handling per HIPAA requirements.

​
Automatic Redaction

Rubric can automatically detect and redact sensitive information: 
client.calls.log(
    transcript=[...],
    
    # Enable PII detection and redaction
    redact_pii=True,
    pii_types=["name", "ssn", "dob", "address", "phone"]
)
The stored transcript will have redactions: 
{
  "speaker": "patient",
  "text": "My name is [REDACTED_NAME] and my date of birth is [REDACTED_DOB].",
  "redactions": [
    {"type": "name", "start": 11, "end": 26},
    {"type": "dob", "start": 51, "end": 63}
  ]
}

​
Next Steps

Voice Overview

Full voice AI evaluation guide

Patient Triage

Triage-specific evaluation