---
title: "Personalize your product's text-to-speech voice for any language: Fine-tuning with Kubeflow Trainer on Red Hat OpenShift AI"
description: "Training an English-centered text-to-speech model for Turkish can produce fluent but hard-to-understand pronunciation for native speakers. This tutorial fine-tunes unsloth/orpheus-3b-0.1-pretrained with about 81,000 Turkish text-audio pairs on Red Hat OpenShift AI using a Kubeflow Trainer TrainJob distributed across two nodes with two A100 GPUs each. LoRA freezes the 3-billion-parameter base and trains 97.3M adapter parameters, while preprocessing converts text to Llama-3 tokenizer tokens and audio to three-layer SNAC codes, masking text labels during scoring. The merged model is evaluated against the base on ten Turkish sentences using Whisper transcription and WER/CER metrics. Mean WER falls from 2.44 to 0.35 and CER from 1.59 to 0.09, though the authors describe the small test as directional."
---

# Personalize your product's text-to-speech voice for any language: Fine-tuning with Kubeflow Trainer on Red Hat OpenShift AI

[Red Hat](https://yomu.fyi/company/red-hat) · Dmytro Hryshchenko, Abhijeet Dhumal · Sep 9, 2026

**Type:** Tutorial

## Summary

Training an English-centered text-to-speech model for Turkish can produce fluent but hard-to-understand pronunciation for native speakers. This tutorial fine-tunes unsloth/orpheus-3b-0.1-pretrained with about 81,000 Turkish text-audio pairs on Red Hat OpenShift AI using a Kubeflow Trainer TrainJob distributed across two nodes with two A100 GPUs each. LoRA freezes the 3-billion-parameter base and trains 97.3M adapter parameters, while preprocessing converts text to Llama-3 tokenizer tokens and audio to three-layer SNAC codes, masking text labels during scoring. The merged model is evaluated against the base on ten Turkish sentences using Whisper transcription and WER/CER metrics. Mean WER falls from 2.44 to 0.35 and CER from 1.59 to 0.09, though the authors describe the small test as directional.

## Context

Most open text-to-speech models are trained primarily on English, so Turkish pronunciation can substitute nearby English sounds for letters including ç, ğ, ı, ö, ş, and ü. The work addresses this language-specific speech quality problem using Turkish paired text and audio data.

## Approach / What changed

Fine-tune the Orpheus-3B pretrained model with LoRA on Red Hat OpenShift AI, orchestrating distributed training through a Kubeflow Trainer TrainJob. The workflow tokenizes text, encodes 24 kHz audio with the three-layer SNAC codec, masks text labels, merges the best adapter checkpoint into the base model, and compares generated speech with the original model using Whisper-based WER and CER measurements.

## Takeaways

- The TrainJob exposes the base model, dataset, epoch count, and LoRA rank as environment variables, so changing the language or training length requires configuration edits rather than training-code changes.
- The tokenizer must match the model; a mismatch can leave training running and the loss falling while the resulting audio remains unusable.
- Evaluation loss is calculated on a 1,000-sample held-out split, while WER and CER come from a separate ten-sentence Turkish dictation test covering registers such as directions, news, weather, and flight announcements.

**Tags:** [GPU](https://yomu.fyi/topic/gpu), [Kubeflow](https://yomu.fyi/topic/kubeflow), [LoRA](https://yomu.fyi/topic/lora), [OpenShift AI](https://yomu.fyi/topic/openshift-ai)

- Source: [Red Hat](https://developers.redhat.com/articles/2026/09/09/text-to-speech-for-any-language-fine-tuning-with-kubeflow-trainer-on-red-hat-openshift-ai)
- Source URL: https://developers.redhat.com/articles/2026/09/09/text-to-speech-for-any-language-fine-tuning-with-kubeflow-trainer-on-red-hat-openshift-ai
- Ingested by Yomu: 2026-09-09T04:00:51.308Z

[Read original post](https://developers.redhat.com/articles/2026/09/09/text-to-speech-for-any-language-fine-tuning-with-kubeflow-trainer-on-red-hat-openshift-ai)