Azerbaijani Text-to-Speech

The key problem are: Text-to-Speech (TTS) Issues for Azerbaijani Content. The core issue is that the generated voice content lacks the natural quality and expressiveness required for professional use.

Pronunciation and Monotone Output

• The output is frequently monotonous, and word stresses are often incorrect.
• Azerbaijani-specific phonemes are inconsistent.
• Testing native Azerbaijani voices (az-AZ-BanuNeural and az-AZ-BabekNeural) showed they also sound monotonous, mispronounce certain sounds (like [ç]), and incorrectly read foreign words.

Handling of Structured Content

• The system incorrectly handles dates, numbers, abbreviations, and acronyms.
• Abbreviations (e.g., "ABŞ," "4-B," "GPT") are often read letter-by-letter or mispronounced.
• Numbers (e.g., "23000" or "4.5 milyard") are not rendered naturally.
• Ordinal suffixes in dates (e.g., "-cı/-ci") cause issues despite programmatic pre-conversion to spoken text.

SSML Limitations

• The effectiveness of SSML tags for prosody, emphasis, and styles is limited due to language/voice constraints.
• Syllable-level stress control is not supported.
• The SSML style="news" being used is not a supported style value and, if invalid, causes the entire tag to be ignored, falling back to neutral speech.
• Azerbaijani is not currently listed as a language that supports broader SSML "say-as" interpret-as modes (like date, number, time, currency).