Speaker embedding dimension mismatch: 512-dim v1/v2 embeddings tagged as v3 crash cdist comparison

[beastoin]

Problem

4 users are generating ~1,197 errors/day from scipy.cdist dimension mismatch in speaker identification:

ValueError: XA and XB must have the same number of columns (i.e. feature dimension.)

Crash site: compare_embeddings() in speaker_embedding.py when comparing 256-dim (v3 model) vs 512-dim (v1/v2 model) embeddings.

Root Cause (verified via Firestore query)

The v2→v3 speaker embedding migration (speaker_sample_migration.py:186-190) has a bug: when a contact has no speech_samples (0 audio files), it sets speech_samples_version=3 but leaves the old 512-dim embedding untouched.

Verified on affected user 0hqtnbQeNR...:

Contact	emb_dim	version	speech_samples	Status
Kat	512	3	0	BUG — v3 tag, old embedding
John P.	512	NOT SET	0	Legacy, never migrated
Chris Bond	256	3	1	Correct — re-extracted
Vince	256	3	1	Correct — re-extracted

The migration path: if not samples: update_version(3); return — bumps version without clearing or re-extracting the old embedding. Same bug exists in v1→v2 path (line 73-77).

Fix

PR #6240:

1. compare_embeddings() returns 2.0 (max distance) on dimension mismatch — crash prevention
2. Migration clears speaker_embedding when no samples exist (both v1→v2 and v2→v3)
3. Hardening: transcribe.py and sync.py skip loading embeddings for contacts without speech_samples

Files to Modify

• backend/utils/stt/speaker_embedding.py — dimension guard in compare_embeddings()
• backend/utils/speaker_sample_migration.py — clear stale embedding in no-samples path
• backend/routers/transcribe.py — skip loading embeddings without speech_samples
• backend/routers/sync.py — same hardening

---

by AI for @beastoin