Speech synthesis for English training with Spanish phrases
Shortlist MOSI Intelligence’s MOSS-TTS-v1.5 and Eleven Multilingual v2 for English training that includes Spanish phrases. Test the transition inside a sentence; separate support for both languages does not establish good mixed-language delivery.
MOSS-TTS documents multilingual and code-switched synthesis, including English and Spanish. ElevenLabs lists English and Spanish for Multilingual v2. These are reasons to test the models, not a comparative pronunciation result.
Compare the controls that matter for the lesson
| Candidate | Officially documented support | What still needs a bilingual reviewer |
|---|---|---|
| MOSS-TTS-v1.5 | Code-switching and Pinyin/IPA controls | Exact phrase reading and natural switches in your chosen voice |
| Eleven Multilingual v2 | English and Spanish in its language list | Within-sentence delivery for your regional pronunciation needs |
MOSS-TTS-v1.5 is a particularly relevant trial when pronunciation corrections must be retained in an open generation pipeline. Keep a short plain-text version and a pronunciation-controlled version of each difficult passage. Compare both before adopting a control globally; a change that fixes one term may disturb the surrounding sentence. ElevenLabs labels Multilingual v2 as a previous-generation option, so this is a focused baseline rather than an exhaustive survey of its current models.
Build a bilingual acceptance set
Use a bilingual reviewer to approve expected readings before synthesis. Include a Spanish name, a short quoted phrase, a full Spanish instruction and a switch back to English. Test punctuation, numbers and abbreviations around the language boundary.
| Test case | What to listen for |
|---|---|
| Spanish phrase within English | Correct pronunciation without an unwanted voice change |
| Spanish proper noun | A recognizable name in the intended regional reading |
| Return to English | No persistent unintended accent shift |
| Instruction with numbers | Correct value and understandable spoken form |
| Repeated course term | The same reading across lessons |
A regional accent preference should be specified by the training team. Do not mark an otherwise valid pronunciation wrong simply because it differs from a reviewer’s accent.
Keep one voice policy across the course
Record the chosen voice or authorized reference, normalization rules and approved term readings. Reuse these settings when lessons are updated. If the model exposes language controls, test their behavior on genuinely mixed text rather than forcing an entire bilingual sentence into one language setting.
Review synthesis before adding music or effects. Those additions can mask pronunciation errors during approval while leaving them audible to a learner using headphones.
When MOSI is the better fit
Prioritize MOSS-TTS-v1.5 when an open synthesis workflow and documented code-switching are useful to your team. Choose the final system on bilingual intelligibility, consistent voice and correction effort. Cloud access, language support and model names should be recorded with the review date, because service and release behavior can change.
This is a proposed evaluation method based on documentation checked on November 1, 2026; it reports no hands-on comparison.
Sources and product documentation checked November 1, 2026. This guide proposes an evaluation workflow and reports no hands-on benchmark results.