Azure Speech TTS: Unnatural stress on Spanish words ending in "-tón" across multiple neural voices

Hello Microsoft Speech Team,

While testing Azure AI Speech for a Spanish language-learning application, I noticed a pronunciation issue affecting several Spanish words ending in "-tón", particularly when they are spoken alone or at the end of a phrase.

Initially, I thought this might be an issue with the accent mark (ó) in the word "ratón". However, further testing suggests that the problem may be related to how Azure Speech handles word stress and sentence-final prosody.

1. Words showing the issue

The following words consistently sound unnatural when synthesized individually:

• ratón
• botón
• montón
• cartón
• bastón

All these words carry lexical stress on the final syllable.

The issue is that the expected stress on the final syllable sounds weak, unclear, or unnatural. In some voices, the effect is more noticeable than in others.

2. Comparison with similar words

To determine whether the problem was related to the accented vowel "ó" or the ending "-ón", I tested several other words:

• perdón
• cordón
• algodón
• canción
• corazón
• camión
• limón

These words sound normal.

This suggests that the problem may be associated specifically with the phonetic sequence "-tón", rather than the accented vowel or the "-ón" ending in general.

3. The pronunciation changes depending on context

An especially interesting observation is that "ratón" sounds correct when followed by another word.

For example:

Text	Observed result
ratón	Unnatural stress
--------	--------
ratón	Unnatural stress
el ratón	Unnatural stress
ratón grande	Normal pronunciation
ratón pequeño	Normal pronunciation
un ratón grande	Normal pronunciation
ratón azul	Normal pronunciation
ratón blanco	Normal pronunciation
ratones	Normal pronunciation

The fact that the same word sounds natural when followed by another word makes me wonder whether the issue is related to sentence-final prosody rather than incorrect recognition of the accent mark.

4. Multiple Spanish neural voices are affected

I initially noticed the issue with:

es-ES-ElviraNeural

However, I subsequently tested other Spanish (Spain) neural voices, as well as voices from Spanish (Mexico).

The same general pronunciation issue was noticeable across multiple voices, although the severity varied depending on the voice.

Some voices sounded relatively close to the expected pronunciation, while others sounded considerably less natural.

This suggests that the issue may not be limited to a single neural voice.

5. Testing environment

I first observed the issue in my application, which uses Azure Speech TTS.

I then reproduced the same general behavior directly in Microsoft Azure Speech Studio, without applying any custom IPA phoneme overrides.

This suggests that the behavior is not specific to my application's text-processing logic.

6. Questions for the Speech team

I would appreciate your guidance on the following:

1. Could this behavior be related to sentence-final prosody or lexical stress realization in Spanish neural TTS?
2. Is there a known limitation affecting words ending in "-tón"?
3. Since the behavior appears across multiple Spanish neural voices, could it be related to a shared pronunciation or prosody-processing component?
4. Is there a way to report this as a speech-quality issue for engineering review without requiring a paid Azure technical support plan?

I understand that SSML phoneme or prosody controls may provide a workaround. However, since the application is intended for Spanish language education, accurate default pronunciation is important.

I would be happy to provide additional test examples, synthesized audio samples, or SSML payloads if needed.

Thank you for your assistance.Hello Microsoft Speech Team,

While testing Azure AI Speech for a Spanish language-learning application, I noticed a pronunciation issue affecting several Spanish words ending in "-tón", particularly when they are spoken alone or at the end of a phrase.

Initially, I thought this might be an issue with the accent mark (ó) in the word "ratón". However, further testing suggests that the problem may be related to how Azure Speech handles word stress and sentence-final prosody.

1. Words showing the issue

The following words consistently sound unnatural when synthesized individually:

• ratón
• botón
• montón
• cartón
• bastón

All these words carry lexical stress on the final syllable.

The issue is that the expected stress on the final syllable sounds weak, unclear, or unnatural. In some voices, the effect is more noticeable than in others.

2. Comparison with similar words

To determine whether the problem was related to the accented vowel "ó" or the ending "-ón", I tested several other words:

• perdón
• cordón
• algodón
• canción
• corazón
• camión
• limón

These words sound normal.

This suggests that the problem may be associated specifically with the phonetic sequence "-tón", rather than the accented vowel or the "-ón" ending in general.

3. The pronunciation changes depending on context

An especially interesting observation is that "ratón" sounds correct when followed by another word.

For example:

Text	Observed result
ratón	Unnatural stress
el ratón	Unnatural stress
ratón grande	Normal pronunciation
ratón pequeño	Normal pronunciation
un ratón grande	Normal pronunciation
ratón azul	Normal pronunciation
ratón blanco	Normal pronunciation
ratones	Normal pronunciation

The fact that the same word sounds natural when followed by another word makes me wonder whether the issue is related to sentence-final prosody rather than incorrect recognition of the accent mark.

4. Multiple Spanish neural voices are affected

I initially noticed the issue with:

es-ES-ElviraNeural

However, I subsequently tested other Spanish (Spain) neural voices, as well as voices from Spanish (Mexico).

The same general pronunciation issue was noticeable across multiple voices, although the severity varied depending on the voice.

Some voices sounded relatively close to the expected pronunciation, while others sounded considerably less natural.

This suggests that the issue may not be limited to a single neural voice.

5. Testing environment

I first observed the issue in my application, which uses Azure Speech TTS.

I then reproduced the same general behavior directly in Microsoft Azure Speech Studio, without applying any custom IPA phoneme overrides.

This suggests that the behavior is not specific to my application's text-processing logic.

6. Questions for the Speech team

I would appreciate your guidance on the following:

1. Could this behavior be related to sentence-final prosody or lexical stress realization in Spanish neural TTS?
2. Is there a known limitation affecting words ending in "-tón"?
3. Since the behavior appears across multiple Spanish neural voices, could it be related to a shared pronunciation or prosody-processing component?
4. Is there a way to report this as a speech-quality issue for engineering review without requiring a paid Azure technical support plan?

I understand that SSML phoneme or prosody controls may provide a workaround. However, since the application is intended for Spanish language education, accurate default pronunciation is important.

I would be happy to provide additional test examples, synthesized audio samples, or SSML payloads if needed.

Thank you for your assistance.