Speaker Attribution from Text
Who said it matters more than what was said: one wrong speaker label can fabricate a customer's intent.
Why this matters
Agents ask leading questions (“so you want 5 lakh, right?”), paraphrase and pitch. If or role assignment attributes those lines to the customer, every downstream stage sees customer “evidence” that the customer never said — and produces a confident, wrong, business-relevant output.
This is the textbook case of : a small upstream error that no downstream component can detect on its own. Speaker attribution therefore needs its own evaluation and, often, text-based checks on top of audio diarization.
Core concept
Technical explanation + simple intuition + example
Linguistic clues to speaker role
Agents and customers talk differently. Agents greet, introduce the company, ask qualification questions, use scripted phrases and say “sir”. Customers state needs (“mujhe… chahiye”), give short answers (“haan”, “five lakh”), object and ask about costs and documents.
No single clue is reliable — customers say “sir” too — but together they give a strong signal.
Error propagation
Speaker error → wrong evidence attribution → wrong intent classification → wrong business decision. Each stage did its job correctly *given its input*, so none of them raises an alarm. The final output looks confident and well-evidenced.
The 10 questions
Every important concept should answer all of these
1What is it?
Assigning each utterance to a speaker role (Agent or Customer), from audio diarization and/or textual cues.
2Why does it exist?
All customer-centric outputs (intent, objections, consent) depend on knowing which words were the customer's.
3What problem does it solve?
It prevents agent speech — questions, paraphrases, pitches — from being treated as customer evidence.
4How does it work internally?
Diarization clusters voices; role mapping uses text cues (questions, scripts, self-reference, need expressions) and alternation constraints; LLMs can check consistency.
5What does it look like on real transcript data?
Mono telephony recordings, overlapping speech and short back-channels cause swapped or merged turns; Hinglish politeness ('sir', 'ji') appears on both sides.
6What can go wrong?
Leading questions attributed to customers; merged turns mixing both speakers; systematic swaps for a whole call; back-channels assigned to the wrong side.
7How do I evaluate it?
Turn-level role accuracy on a labelled sample, and — more importantly — accuracy on evidence-bearing turns, sliced by audio quality and agent.
8When should I use it?
Always verify speaker roles for any customer-intent, consent or complaint extraction.
9When should I NOT use it?
Tasks that describe the whole conversation neutrally (topic tagging) are less sensitive — but still check.
10How does it affect downstream systems?
Wrong attribution fabricates evidence, inflates 'interested' rates, misattributes consent in compliance, and corrupts agent-performance metrics.
Interactive lab: Who said it? + error-propagation simulator
Assign speakers to unlabelled lines and see the linguistic cues behind each decision. Then flip speaker labels in a call and watch the error cascade through evidence, intent and the business decision.
Try it yourself
Conceptual (A), applied (B) and engineering (C) exercises
Graded by a transparent keyword rubric (not an LLM) — it checks for the key ideas, then shows a model answer for self-assessment.
Real-world application
A lender's “interested lead” rate jumped 11% after a telephony vendor switch. Nobody changed the model. Per-stage evaluation found that the new mono recordings degraded diarization: agent confirmations (“so you'll take 5 lakh, right?”) were attributed to customers in 8% of calls. Sales wasted thousands of callbacks. The fix combined stereo recording, a text-based role checker that flags customer-attributed interrogatives, and a rule that leading questions alone can't support “interested”.
Common mistakes
Diarization errors are common on mono, noisy, overlapping audio.
Fix: Evaluate diarization on a sample and add text-based consistency checks.
A leading question + 'hmm' is not a customer statement of intent.
Fix: Require evidence from customer turns that are not interrogatives; lower confidence when evidence is only a 'yes' to a leading question.
It can't reveal that speaker errors caused the intent errors.
Fix: Measure speaker attribution separately and on evidence-bearing turns.
Engineering decision
How much effort should go into speaker attribution?
- Outputs concern the customer's own statements (intent, consent, complaints)
- Recordings are mono or noisy
- Decisions are costly (sales routing, compliance)
- You have reliable stereo channels per party (then verify, but don't over-engineer)
Quiz
Pass mark 60%. Missed questions go to your mistake log and spaced review.