> For the complete documentation index, see [llms.txt](https://documentation.proto.cx/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://documentation.proto.cx/docs/developers/local-languages.md).

# Local languages

Proto supports low-resource languages in Africa and Asia. This section covers our latest performance benchmarks for priority local languages. It also explains how to prepare the data that trains our local language models.

***

## Model performance

Measured performance of the deployed ASR (speech-to-text) and TTS (text-to-speech) models for the local languages, against Proto's target thresholds.

### ASR (speech-to-text)

| Language    | Accuracy rate (target ≥ 90%) | Word error rate (target ≤ 10%) | Fallback rate (target < 5%) | Response speed (target < 1s) |
| ----------- | ---------------------------- | ------------------------------ | --------------------------- | ---------------------------- |
| Filipino    | 90.66%                       | 9.34%                          | 4.21%                       | 1.74s                        |
| Cebuano     | 87.74%                       | 12.26%                         | 6.36%                       | 1.76s                        |
| Kinyarwanda | 93.73%                       | 6.27%                          | 4.93%                       | 1.75s                        |
| Oshiwambo   | 70.22%                       | 29.78%                         | 8.34%                       | 1.77s                        |

In development — baselines measured before fine-tuning, expected to improve:

| Language | Accuracy rate (target ≥ 90%) | Word error rate (target ≤ 10%) | Fallback rate (target < 5%) | Response speed (target < 1s) |
| -------- | ---------------------------- | ------------------------------ | --------------------------- | ---------------------------- |
| Urdu     | 79.4%                        | 20.6%                          | Not yet measured            | Not yet measured             |
| Punjabi  | 51.5%                        | 48.5%                          | Not yet measured            | Not yet measured             |

Word error rate broken down by audio condition:

| Language    | Clean WER (target ≤ 15%) | Noisy WER (target ≤ 35%) |
| ----------- | ------------------------ | ------------------------ |
| Filipino    | 9.34%                    | 15.25%                   |
| Cebuano     | 15.26%                   | 25.85%                   |
| Kinyarwanda | 10.27%                   | 16.52%                   |
| Oshiwambo   | Retesting                | Retesting                |

### TTS (text-to-speech)

| Language    | Mean opinion score (target 4.5/5) | Fallback rate (target < 5%) | Response speed (target < 1s) |
| ----------- | --------------------------------- | --------------------------- | ---------------------------- |
| Filipino    | 4.0                               | 17.32%                      | 1.25s                        |
| Cebuano     | 3.8                               | 22.58%                      | 1.53s                        |
| Kinyarwanda | 4.0                               | 31.46%                      | 2.26s                        |
| Oshiwambo   | 3.6                               | 31.24%                      | 2.37s                        |

In development:

| Language | Mean opinion score (target 4.5/5) | Fallback rate (target < 5%) | Response speed (target < 1s) |
| -------- | --------------------------------- | --------------------------- | ---------------------------- |
| Urdu     | 4.15                              | Not yet measured            | Not yet measured             |

{% hint style="info" %}
Mean opinion score (MOS) rates the naturalness of synthesised speech on a 1–5 scale (5 excellent, 1 bad). Tests were run on an Azure Standard NC4as T4 v3 instance (28 GB RAM, NVIDIA Tesla T4 GPU); the ASR test set used audio samples of 5–10 seconds.
{% endhint %}

***

### In this section

{% content-ref url="/pages/Xt2hGzGPPi7WqJeWLcAS" %}
[Language training](/docs/developers/local-languages/language-training.md)
{% endcontent-ref %}

{% content-ref url="/pages/LffWZfYXvAgtNRag4bLY" %}
[Voice data collection](/docs/developers/local-languages/language-training/voice-data-collection.md)
{% endcontent-ref %}

{% content-ref url="/pages/dCS14f5rFaYHkrr9eCp6" %}
[Text data preparation](/docs/developers/local-languages/language-training/text-data-preparation.md)
{% endcontent-ref %}