sl-SI - Petra and Rok - Incorrect lexical stress and abbreviation handling

TITLE: sl-SI PetraNeural / RokNeural – Systematic lexical stress bug – NoSpeechResources auto-redirect – Radio Slovenc DESCRIPTION – COPY TO AZURE SUPPORT: Hello team, Following Q&A 5951967 answered by Christos Panagiotidis. We cannot provide Request ID because Speech Studio auto-redirects us to Foundry: https://ai.azure.com/...?Trigger=AutoRedirect_NoSpeechResources We have no CognitiveServices resource yet (we use Petra via third-party radio automation that uses Azure TTS in backend). Foundry page is stuck on "Loading..." even after clicking "Switch to the old look". We therefore report via this ticket without Request ID, as advised in docs for language bugs. VOICES: sl-SI-PetraNeural, sl-SI-RokNeural REGION: westeurope (tried both) SSML that fails (minimal, valid): <speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis" xml:lang="sl-SI"><voice name="sl-SI-PetraNeural">Gasilci so na poti domov. Pot je zaprta zaradi lesa. Požarna ogroženost je velika. NMP v PGD Krško.</voice></speak> Issues: 1. Lexical stress: gaSILci (should GA:), poTI (should PO:), doMOV, leSA, poZARna. Slovenian has fixed initial stress, model puts it on second syllable. IPA in attached lexicon. 2. Abbreviations NMP, PGD, UKC, ZD read as words, not characters. say-as characters workaround works but need default for SI emergency. 3. L-vocalization: pihal should be [pi:hau] not hard L. Attachments: sl-SI-Radio-Slovenc-lexicon.pls with 30 verified IPA forms. We are Radio Slovenc – willing to provide 500-word lexicon and 10h reference audio to fix model permanently for all Slovenian users. Please do not close as "needs Request ID" – we are blocked by AutoRedirect_NoSpeechResources bug. Q&A cannot take private files. Contact: Bozidar Brajer, Radio Slovenc, Krmelj, Slovenia. --- Q&A REPLY TO POST AT https://learn.microsoft.com/en-us/answers/questions/5951967/ : Thank you Christos for confirmation and SSML fix. say-as characters works for NMP/PGD. We tried to get Request ID via Speech Studio but we are blocked by AutoRedirect_NoSpeechResources – Foundry redirects to https://ai.azure.com/... and stays on Loading even after "Switch to the old look". We have no Speech resource yet (we use Petra via radio automation backend). We opened formal support ticket with lexicon attached (sl-SI-Radio-Slovenc-lexicon.pls) and detailed SSML without Request ID as we are blocked. Please advise if you can forward to acoustic-model team without Request ID, or enable trial so we can generate one. Bozidar, Radio Slovenc

Thank you Christos for confirming! We understood the SSML fix with </speak> and <voice> and say-as characters – works for NMP/PGD, thank you.

We agree hyphenation is not a solution. We will open a formal Azure Speech Support Request with voice names sl-SI-PetraNeural / RokNeural, region westeurope, exact SSML, Request IDs and audio samples as you suggested, and propose a custom lexicon with IPA for bulletin vocabulary.

Radio Slovenc, Božidar

The pronunciation examples are reproducible issues worth reporting, but the shown SSML is not valid: it opens <speak> and closes </govor>, and Azure Speech requires a <voice> element inside <speak>.

Test each voice with a minimal document: <speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis" xml:lang="sl-SI"><voice name="sl-SI-PetraNeural">Gasilci so na poti. <say-as interpret-as="characters">NMP</say-as> v <say-as interpret-as="characters">PGD</say-as> Krško.</voice></speak>. The characters interpretation is supported across text-to-speech locales and makes abbreviation handling deterministic.

For incorrect lexical stress, Petra and Rok support <phoneme alphabet="ipa"> and custom lexicons. Add verified Slovenian IPA with primary-stress marks for recurring words, or use a lexicon for the bulletin vocabulary; do not rely on hyphenation as a guaranteed pronunciation control.

If the same words remain wrong without overrides, open an Azure Speech support request and attach the voice name, region, exact SSML, synthesis request IDs, and paired original/corrected audio. Q&A cannot forward samples privately to the acoustic-model team.