Adapting F5-TTS Model to Kurdish Sorani: Diffusion-based Speech Synthesis with a Specialized Dataset

1. INTRODUCTION

Text-to-speech (TTS), also known as speech synthesis, is a technology that converts written words into spoken language. It plays an important role in modern technology. Most artificial intelligence chatbots, smart devices, and even organizations use such systems to provide fast, more realistic, and enjoyable service. For example, by converting written text into natural-sounding speech, this technology is mainly utilized in human-computer interaction systems to make them more user-friendly [1]–[3]. An advanced TTS system should not only read what has been written, but also be able to maintain emotion, rhythm, punctuation, and many other properties to produce speech that sounds like a real human.

However, achieving these capabilities is not equally accessible across all languages. For widely spoken languages such as English, Spanish, and Arabic, TTS technology has been improved significantly [4], while low-resource languages are neglected and outdated [5]. In the case of Kurdish, while recent studies [6]–[8] have made noticeable progress, they often show limitations in fluency and emotional expression, as indicated by their reported mean opinion score (MOS) of 4.1 [8] and 3.9 [7]. The two major reasons for this gap are: unavailability of a high-quality, domain-diverse dataset, and the use of outdated TTS models that are not suitable for low-resource languages [5].

The Kurdish Sorani script is similar to that of Arabic, Persian, and other related languages. These languages share some phonetic consistency [9], [10]. Most words in Kurdish Sorani script are pronounced as they are written. However, this phonetic transparency does not eliminate challenges faced when building a TTS dataset. For example, during natural speech or when expressing emotions, speakers often skip or merge letters, alter pronunciation, or blend sounds [11]. In addition, the Kurdish language also lacks a standardized punctuation convention [11], [12], which makes it more difficult to map Kurdish script to speech. These issues introduce more complexity to the process of creating a clean and usable dataset for TTS models.

Given these inherent challenges in Kurdish TTS development, this work presents the first fine-tuned F5-TTS model for the Kurdish Sorani language. F5-TTS, which stands for A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching, is a recently developed open-source model that leverages flow-matching and diffusion-based techniques to generate natural and emotional speech without the need for extensive training data [13]. The process of fine-tuning this model began with the collection of 10.11 h of high-quality speech from professional TV recordings, which were then segmented into over 4,800 carefully curated samples covering diverse topics and speech styles. Each sample was manually reviewed for transcription accuracy and balanced duration.

The following are key contributions of this study:

The rest of this paper begins with a review of related studies on speech synthesis for low-resource languages and prior efforts in Kurdish TTS. Next, the dataset design is discussed, including how the speech data was collected, segmented, and transcribed. This is followed by an explanation of the F5-TTS model architecture and the fine-tuning process. Finally, the evaluation results are presented, and the implications of the findings are discussed, with concluding remarks and directions for future research.

2. RELATED WORK

TTS technology has seen fast advancements in recent years. However, for low-resource languages such as Kurdish, progress has been slower due to limited attention and support in the field of speech synthesis. This section provides an overview of traditional and neural speech synthesis methods, with a focus on approaches adapted to low-resource settings, and mentions previous efforts related to Kurdish TTS systems.

3. DATASET DESIGN

Building a speech dataset from scratch is a gradual and careful process in which every single detail matters [20], [29]. You must care about every single detail. In the case of TTS, every single word matters because the quality and the diversity of the training data will undoubtedly impact the performance of the resulting model.

3.2. Data Segmentation

After data collection, the dataset passed several preparation stages, with segmentation and transcription being the most time-consuming steps. In data segmentation, audio files were divided into manageable clips while ensuring the following:

• Proper logical sentence boundaries

• Natural starting and ending tone.

• A balanced distribution of clip durations.

To achieve a balanced dataset, segments were distributed approximately equally across different duration ranges (e.g., 2–4 s, 4–6 s, 6–8 s, etc.), as shown in Fig. 1. This prevented the model from overfitting to specific speech length ranges. During this process, Ocenaudio software was utilized, as it allows the manual selection and export of clean segments.

Fig. 1. Distribution of dataset samples.

3.4. Dataset Statistics

The final version of the dataset was 10.11 h long in 4,856 samples. Each sample was stored in WAV format. The dataset vocabulary had 2,567 unique tokens. Detailed statistics are provided in Table 1.

TABLE 1: Training dataset statistics

demonstrates how samples are distributed evenly based on duration, which helps the model generalize across different speech lengths.

Fig. 2. Proposed model workflow.

4. F5-TTS MODEL FINE-TUNING

Fine-tuning TTS models for low-resource languages like Kurdish Sorani can be challenging, primarily due to the lack of training data and limited computational resources. This section describes the steps taken to fine-tune the F5-TTS model for Kurdish Sorani using a specialized speech dataset. It includes an overview of the F5-TTS architecture, the data preparation and training environment, and a description of the fine-tuning process. An overview of the full training pipeline is presented in Fig. 2.

4.1. Overview of F5-TTS Architecture

Recent advancements in TTS technologies have improved the naturalness and flexibility of synthesized speech. Previous TTS systems were primarily based on auto-regressive (AR) models, which generate speech one token at a time, with each step depending on the last. Although models such as variational inference with adversarial learning for end-to-end text-to-speech and Tacotron2 [19] achieved high-quality results, they still suffer from inference latency and errors during sequential generation.

F5-TTS is a fully non-autoregressive (NAR) TTS system, meaning it generates all audio tokens in parallel rather than sequentially, that produces high-quality speech using a flow matching framework, a generative modeling technique that learns to map noise into realistic data distributions [30], combined with a diffusion transformer (DiT) backbone, a transformer-based architecture tailored for diffusion models [31]. NAR approaches, particularly those based on diffusion-based models, which iteratively denoise random noise to produce realistic data [32], and flow matching techniques, allow for faster and more robust synthesis.

Unlike traditional TTS systems that rely on phoneme alignment, explicit duration models, or complex text encoders, F5-TTS presents a simplified architecture. It can directly learn how to speak from text characters and noisy speech. The model architecture can be described in five main phases, as shown below:

• Input Processing: The input text is turned into a sequence of characters, then padded with filler tokens to match the length of target speech frames. This enables the model to learn implicit alignment during training, which leads to a more natural rhythm.

• Feature Refinement: Input text features are refined using ConvNeXt blocks, a modern convolutional neural network design that improves feature extraction efficiency and alignment with speech frames.

• Denoising Process: A DiT gradually denoises a sampled noisy speech signal, conditioned on the refined text representation. It turns the noise into clear speech based on input text.

• Training Objective: Training optimizes a flow-matching loss combined with a text-guided speech infilling task.

• Inference Strategy: Since diffusion models are typically slow due to step-by-step denoising, F5-TTS employs a sway sampling technique, an accelerated sampling method that reduces generation time while maintaining naturalness.

An overview of F5-TTS training (left) and inference (right) flow is illustrated in Fig. 3.

Fig. 3. F5-text-to-speech model architecture [13].

5. RESULTS DISCUSSION AND EVALUATION

The loss curve shown in Fig. 4 proves that the model has achieved excellent convergence and effective learning throughout the training process. To further assess the performance of the fine-tuned model, three evaluation metrics were used: CER, WER, and MOS. CER and WER are objective metrics that measure transcription accuracy at the character and word level, while MOS is a 1–5 scale for subjective rating of naturalness.

For objective metrics, the synthesized speech was transcribed using a pretrained Whisper Kurdish Sorani Automatic Speech Recognition (ASR) model (“PawanKrd/asr-large-ckb”), and the transcriptions were then compared to the original text. In contrast, the MOS metric was obtained based on human listener ratings, as the naturalness of speech may not be fully captured by objective metrics alone.

The results of CER and WER for three different samples are presented in Table 5. In both metrics, lower average scores indicate better synthesis quality. The average WER was 20.37%, and the average CER was 4.3%. These results demonstrate the model’s strong transcription accuracy, particularly at the character level. The relatively high WER value of 20.37% may be due to limitations in the ASR model for Kurdish Sorani rather than actual quality issues, as demonstrated by the low CER and high MOS scores.

TABLE 5: Word error rate (WER) and character error rate (CER) for synthesized Kurdish speech samples

To complement these objective measures with human perception assessments, a subjective evaluation was conducted using the MOS metric. Synthesized speech samples were shared online with more than 50 different native Kurdish Sorani speakers of different ages and genders. Each listener was asked to rate the naturalness of the same set of synthesized speech clips on a five-point scale ranging from 1 (Bad) to 5 (Excellent). The average score achieved in this metric was 4.72 out of 5, implying that the generated speech was perceived as highly natural. While this result is promising, it is important to remember that the number of listeners was limited, and no statistical significance testing was applied. Future evaluations with larger participant groups and formal statistical analysis would lead to more reliable results.

At the same time, these findings were obtained using clean single-speaker data in controlled settings. Several factors that may impact system performance in real-world scenarios, including background noise, overlapping speech, as well as speaker and dialect variability. Handling these challenges may require additional techniques and extended datasets, which can be explored in future research directions.

In comparison to older TTS approaches for low-resource languages such as transfer learning, the results of our diffusion-based model show significant improvements. Table 6 provides a comparison of various methods applied to low-resource languages, including different architectures and learning strategies. Notably, while prior Kurdish TTS studies have reported their MOS scores, none have evaluated their systems using objective evaluation metrics, such as CER or WER. This work is the first to apply these objective measures to Kurdish Sorani TTS for a more precise evaluation.

TABLE 6: Comparison of text-to-speech systems for low-resource languages

Beyond the technical achievements, this work has significant cultural and linguistic importance. Developing high-quality TTS systems for underrepresented Kurdish Sorani can support digital inclusion, education, and even cultural preservation. It ensures the language is not left behind in the advancements of TTS technology. The fine-tuned model becomes a core module in many applications such as audiobooks, e-learning materials, and automated dubbing tools. In addition, building computational resources for Kurdish helps document and standardize the language, which faces dialectal diversity and limited digital resources. For this reason, the benefits of this research go beyond speech synthesis only; it can serve both practical applications and the preservation of Kurdish language identity.

6. CONCLUSION AND FUTURE WORKS

In this study, the first F5-TTS model was successfully fine-tuned for Kurdish Sorani, a low-resource language that has received less attention in the field of speech synthesis. The process began by creating a 10.11-h speech dataset. The work focused mostly on balancing segments, topic variety, emotions, transcription accuracy, punctuation, and other details that optimized the dataset quality. Using this high-quality dataset, the open-source F5-TTS model was fine-tuned in a well-optimized environment and configuration settings. Evaluations showed that the system learned to synthesize natural, emotional, and human-like Kurdish speech.

Compared to previous Kurdish TTS works, this study indicates that modern diffusion-based models like F5-TTS can improve the naturalness and reliability of speech synthesis for low-resource languages without requiring a massive amount of training data or using a pre-trained English-based architecture. The results show that a carefully constructed dataset can overcome the typical limitations of low-resource TTS systems.

While the fine-tuned model provides good results, there are still some areas for further improvement. One limitation is the use of a single-speaker dataset, which may reduce generalizability to different voices or dialects. Using a multi-speaker dataset that includes various emotions and accents would enable additional training to improve this fine-tuned model. This can improve the model’s generalization for different speaking styles. In addition, real-world deployment introduces challenges such as background noise, overlapping speech, and speaker variability, which may reduce accuracy. Addressing these conditions in future work will improve the system’s robustness and support its integration into practical Kurdish applications such as audiobooks, dubbing, and other voice-based systems.