title	Sunbird AI Pipeline
emoji	🐦
colorFrom	green
colorTo	yellow
sdk	gradio
sdk_version	6.14.0
app_file	app.py
pinned	false

Sunbird AI Pipeline

A Gradio web application that accepts text or audio... (rest of your existing README stays the same)

Sunbird AI Pipeline

A Gradio web application that accepts text or audio, then runs it through a four-step AI pipeline — entirely powered by Sunbird AI — to produce a translated, spoken summary in a Ugandan local language.

---

Project description

This app lets a user provide either typed/pasted text or an uploaded audio file. If audio is supplied, it is first transcribed to text using Sunbird's Speech-to-Text API. The text (typed or transcribed) is then summarised using the Sunflower LLM, and the summary is translated into a chosen Ugandan local language — Luganda, Runyankole, Ateso, Lugbara, or Acholi — using the same LLM. Finally, the translated summary is synthesised as a playable audio clip via Sunbird's Text-to-Speech API. All four intermediate outputs (transcript, summary, translated summary, and audio player) are displayed directly in the UI, so the user can inspect every stage of the pipeline.

---

Architecture overview

User input
  │
  ├─ Text (typed / pasted) ──────────────────────────────┐
  │                                                       │
  └─ Audio file (.mp3 / .wav / .ogg / .m4a / .aac)       │
         │                                                │
         ▼  [audio path only]                             │
  ┌──────────────────────┐                                │
  │  Sunbird STT         │                                │
  │  POST /tasks/stt     │                                │
  │  multipart `audio`   │                                │
  └────────┬─────────────┘                                │
           │ transcript text                              │
           └──────────────────────────────────────────► ──┘
                                                          │
                                                          ▼
                                        ┌─────────────────────────────┐
                                        │  Sunflower LLM — summarise  │
                                        │  POST /tasks/sunflower_     │
                                        │        simple               │
                                        │  form-encoded `instruction` │
                                        └───────────────┬─────────────┘
                                                        │ summary
                                                        ▼
                                        ┌─────────────────────────────┐
                                        │  Sunflower LLM — translate  │
                                        │  POST /tasks/sunflower_     │
                                        │        simple               │
                                        │  form-encoded `instruction` │
                                        └───────────────┬─────────────┘
                                                        │ translated summary
                                                        ▼
                                        ┌─────────────────────────────┐
                                        │  Sunbird TTS                │
                                        │  POST /tasks/modal/tts      │
                                        │  JSON `text` + `speaker_id` │
                                        └───────────────┬─────────────┘
                                                        │ signed audio URL
                                                        ▼
                                                  Audio player

Step	Endpoint	Format
Speech-to-Text	POST /tasks/stt	multipart/form-data, field: audio
Summarise	POST /tasks/sunflower_simple	application/x-www-form-urlencoded, field: instruction
Translate	POST /tasks/sunflower_simple	same endpoint, different instruction prompt
Text-to-Speech	POST /tasks/modal/tts	application/json, fields: text, speaker_id

---

Local setup

# 1. Clone the repository
git clone https://github.com/GeorgeLukaanya/internship-assessment.git
cd internship-assessment

# 2. Create and activate a virtual environment
python -m venv venv
source venv/bin/activate          # Linux / macOS
# venv\Scripts\activate.bat       # Windows

# 3. Install dependencies
pip install -r requirements.txt

# 4. Configure environment variables
cp .env.example .env
# Edit .env and replace the placeholder with your real Sunbird AI API token

# 5. Run the app
python app.py
# Opens at http://127.0.0.1:7860

---

Environment variables

Variable	Required	Description
SUNBIRD_API_TOKEN	Yes	Bearer token for all Sunbird AI API calls. Get one at api.sunbird.ai.

Copy .env.example to .env and paste your token. The .env file is listed in .gitignore — never commit it.

# .env.example
SUNBIRD_API_TOKEN=your_token_here

---

Usage

Text input

1. Open the app (http://127.0.0.1:7860 locally, or the deployed link below).
2. Select the Text input tab.
3. Pick a Target language from the dropdown (Luganda, Runyankole, Ateso, Lugbara, or Acholi).
4. Paste or type your text into the text box.
5. Click Summarise & Translate.
6. After 3–6 minutes the results appear:
   • Summary — concise English summary from the Sunflower LLM
   • Translated summary — the summary in the chosen language
   • Synthesised speech — an inline audio player for the translated summary

Audio upload

1. Select the Audio upload tab.
2. Pick a Target language.
3. Upload an audio file (MP3, WAV, OGG, M4A, or AAC). Files longer than 5 minutes are rejected immediately with an error message.
4. Click Transcribe, Summarise & Translate.
5. After 3–6 minutes the results appear:
   • Transcript — the text returned by Sunbird STT
   • Summary, Translated summary, and Synthesised speech as above

The Gradio button shows a spinner while the pipeline runs. Each run makes three sequential API calls (two LLM + one TTS), which accounts for the latency.

---

Deployed link

https://huggingface.co/spaces/musheijaa/sunbird-ai-pipeline

The app is hosted on Hugging Face Spaces (Gradio SDK). The SUNBIRD_API_TOKEN is stored as a Space secret and is never exposed in the repository.

---

Known limitations

• Slow inference — The Sunflower LLM is called twice (summarise + translate) and TTS once per run. Total latency is typically 3–6 minutes; this reflects the Sunbird API's inference speed, not the application itself.
• 5-minute audio cap — Audio longer than 300 seconds is rejected client-side before any API call is made.
• Supported languages — Luganda, Runyankole, Ateso, Lugbara, Acholi. Swahili has a TTS voice (speaker_id 246) but is not a Ugandan local language and is excluded from the picker.
• English source text recommended — The Sunflower LLM summarises most reliably from English. Quality may vary for other input languages.
• Free-tier account — The SUNBIRD_API_TOKEN is a free-tier token; rate limits or quotas imposed by Sunbird AI may apply.
• Audio format detection — Duration checking relies on mutagen. Uncommon or DRM-protected formats may bypass the local check; the Sunbird API enforces its own 100 MB / 10-minute limits as a fallback.