Skip to content

COMPARISON.md

Latest commit

 

History

History
85 lines (61 loc) · 3.21 KB

COMPARISON.md

File metadata and controls

85 lines (61 loc) · 3.21 KB

CallShield vs Traditional STT+LLM Pipeline

Architecture Comparison

Traditional Approach: Speech-to-Text + LLM

flowchart LR
    Audio[Audio] -->|"API Call #1"| STT["STT Service<br/>(Whisper/Deepgram)"]
    STT --> Transcript[Text Transcript]
    Transcript -->|"API Call #2"| LLM[LLM Analysis]
    LLM --> Result[Result]
Loading

CallShield Approach: Voxtral Native Audio

flowchart LR
    Audio[Audio] -->|"API Call #1 (only)"| Voxtral["Voxtral Mini<br/>(native audio reasoning)"]
    Voxtral --> Result[Result]
Loading

Feature Comparison

Aspect Traditional (STT + LLM) CallShield (Voxtral Native)
API calls per analysis 2 (transcribe + analyze) 1 (native audio)
Latency ~5-8 seconds ~2-4 seconds
Vocal cue detection Lost in transcription Preserved (tone, stress, urgency)
Robocall/IVR detection Text patterns only Audio + text signals
Call-center background noise Not detectable Detected
TTS/synthetic voice artifacts Not detectable Detected
Scripted speech cadence Not detectable Detected
Cost per analysis 2× API pricing 1× API pricing
Failure modes STT errors propagate to LLM Single inference, no cascading errors
Privacy exposure Transcript stored/transmitted Audio processed in-memory only

Why Native Audio Matters for Scam Detection

1. Vocal Cues Are Critical Evidence

Phone scammers exhibit distinct vocal patterns that text transcripts cannot capture:

  • Aggressive or threatening tone — IRS/police impersonation calls use intimidating delivery
  • Scripted, rehearsed cadence — Call-center scammers read from scripts with unnatural pacing
  • Background noise — Call centers produce distinctive ambient noise patterns
  • TTS artifacts — Robocalls use text-to-speech with detectable synthetic qualities

2. Single-Pass Efficiency

Traditional pipelines require two serial API calls:

  1. Transcribe audio to text (3-5 seconds)
  2. Analyze transcript for scam patterns (2-3 seconds)

Voxtral processes audio directly in one pass, cutting latency nearly in half.

3. No Information Loss

Transcription is a lossy compression of audio. Even high-quality STT loses:

  • Emotional tone and urgency level
  • Speaker confidence and hesitation patterns
  • Environmental audio context
  • Synthetic voice characteristics

Voxtral's multimodal architecture reasons over the full audio signal.

Real-World Results

Tested against 5 real robocall recordings from the FTC Robocall Audio Dataset:

Sample Voxtral Score Traditional Would Miss
SSN suspension threat 0.70 Robocall IVR pattern (audio-only signal)
SSA legal threat 0.85 Aggressive synthetic voice tone
Fake Amazon charge 0.65 Call-center background noise
Vehicle warranty 0.60 Pre-recorded message cadence
Medicare advisor 0.40 Subtle scripted delivery

All 5 correctly detected — including audio-only signals that a text pipeline would miss entirely.