bug(tts): Hebrew word merging makes speech unintelligible

[shaypal5]

Problem

Azure and Google TTS engines merge adjacent Hebrew words into single non-existent words, making most generated speech unintelligible to native speakers.

Source: Listening test, 2026-05-03 (IT + NEG clips).

Examples (from sp_it_a_0001_00)

Expected	Rendered	Meaning lost
"היי, חשבתי" → "hey, cha-shav-ti"	"ha-ya-cha-shav-ti" (one long word)	"hi" merged with "I thought"
"אגב ניסיתי" → "ah-gav nee-see-ti"	"ag-va-nee-see-tee"	"by the way" merged with "I tried"
"הייתי בדרך הביתה" → "ha-yee-ti ba-de-rech ha-bye-tah"	"kee-nee-tee ba-de-rech ha-bee-tah-leh"	Multiple words merged + mispronounced

This happens throughout all generated clips, not just isolated cases.

Likely causes

1. Missing or insufficient whitespace/pause hints in SSML. Azure he-IL may need explicit <break> tags between words to prevent liaison.
2. Nikud (vowel diacritics) absence. Without nikud, the TTS engine guesses word boundaries and vowelization, often incorrectly.
3. Script generation producing unnormalized text. Stage 1b gender disambiguation may not be inserting enough structural cues.

Possible mitigations

• Insert <break time="50ms"/> between words in SSML (at the renderer level)
• Add nikud to high-error words via the normalization lexicon (Stage 1b)
• Investigate if Google Chirp 3 HD handles word boundaries better than Azure
• Test with explicit phoneme tags (<phoneme alphabet="ipa">) for problematic words

Impact

P0 — Blocker. If the speech is unintelligible, no downstream processing (labels, augmentation, training) has value.

[shaypal5]

Reopened from PR #86 (the #83 fix).

PR #70 was the chosen mitigation for this issue but introduced a 6× Whisper WER regression (#83) and a 63% scene-duration regression (#81) by inserting <break time="50ms"/> between every word pair. PR #86 reverts #70 to restore Whisper baseline; the original word-merging unintelligibility this issue describes is therefore back as an open problem.

Alternative mitigations from this issue's body that were not tried before #70 landed and should be evaluated before re-landing any inter-word-break approach:

1. Add nikud (vowel diacritics) to high-error words via the Stage 1b normalization lexicon. Targets the actual cause (missing vowel disambiguation) rather than treating every word boundary as ambiguous.
2. Explicit <phoneme alphabet="ipa"> tags for known-problematic words, scoped to the specific failures rather than applied globally.
3. Test Google Chirp 3 HD as primary for he-IL — Azure he-IL appears to be the engine that needs the most help; Chirp may handle Hebrew word boundaries differently.
4. If breaks are revisited — try shorter durations (10–20 ms) AND only at boundaries where merging is empirically observed (not every word). Note that the prosody-reset side-effect of <break> may persist even at low durations, so a listening test is required regardless of duration.

Any new attempt should be validated with both:

• A native-speaker A/B listening test (the original failure mode is invisible to ASR metrics).
• A Whisper WER regression check on sp_neu_a_0001 (the prior failure mode is invisible to a listener-only test).