Speaker sample text check drops trimmed samples

[beastoin]

Valid speaker samples are being dropped with text_mismatch because the sample audio is a trimmed window that can span multiple same-speaker segments while the expected text only reflects a single segment. This blocks sample storage and speaker embeddings for affected users.

Current Behavior

• verify_and_transcribe_sample compares transcript against segment text with symmetric similarity.
• Trimmed samples that should be valid fail when expected text is longer or spans merged segments.
• Valid samples are dropped due to text_mismatch.

Expected Behavior

Use a language-agnostic containment check so the transcript can be validated as included in the expected text.

Affected Areas

File	Line	Description
backend/utils/speaker_sample.py	66	Text mismatch check uses symmetric similarity
backend/utils/text_utils.py	1	Only similarity helper exists (no containment helper)

Solution

containment = compute_text_containment(transcript, expected_text)
if containment < MIN_CONTAINMENT:
    return transcript, False, f"text_mismatch: containment={containment:.2f}"

Files to Modify

• backend/utils/text_utils.py
• backend/utils/speaker_sample.py
• backend/tests/unit/test_speaker_sample.py
• backend/tests/unit/test_text_containment.py
• backend/test.sh

Impact

Low — adds a containment check and tests; existing quality checks remain.

---

by AI for @beastoin

[beastoin]

Adding logs for reference:

Sample failed quality check: text_mismatch: similarity=0.07
Sample failed quality check: text_mismatch: similarity=0.09
Sample failed quality check: text_mismatch: similarity=0.20
Sample failed quality check: text_mismatch: similarity=0.10
Sample failed quality check: text_mismatch: similarity=0.04
Sample failed quality check: text_mismatch: similarity=0.21
Sample failed quality check: text_mismatch: similarity=0.11
Sample failed quality check: text_mismatch: similarity=0.35
Sample failed quality check: text_mismatch: similarity=0.47
Sample failed quality check: text_mismatch: similarity=0.07
Sample failed quality check: text_mismatch: similarity=0.19
Sample failed quality check: text_mismatch: similarity=0.06
Sample failed quality check: text_mismatch: similarity=0.29

---

by AI for @beastoin