AI Toolbox – User Tutorial

Welcome to the AI Toolbox

The AI Toolbox is Helen Keller's browser-based application that uses artificial intelligence to transcribe audio recordings and translate documents. It is built for M&E specialists and program staff who work with qualitative data — interviews, focus group discussions, and field reports — across Helen Keller's country programs and multiple languages.

The toolbox currently offers two services:

Interface Overview

When you first open the AI Toolbox, you will see the main screen with the following areas:

Figure 1: Full AI Toolbox interface — Audio Analysis mode.

Key Areas

1. Mode Switcher — Toggle between Audio Analysis and Document Translation (Beta) at the top.
2. Input Options Card — Where you configure language, recording type, prompt, and upload your file. The AI model is automatically selected.
3. Process Log & Result Card — The dark terminal-style log shows real-time status messages. Results appear in the text area below.
4. Info Button (i) — The blue circle in the top-right of the Input Options card opens the Quick Start Guide.

Figure 2: Mode switcher — Audio Analysis active.

Audio  Service 1 — Audio Transcription & Translation

This is the primary service. Use it when you have an audio recording (interview, FGD, or IDI) and need a written transcript in the original language — with the option to translate it afterward.

1

Select Your Inputs

Fill in the three required dropdowns in the Input Options card:

• Primary Language — The main language spoken in the audio (e.g., Nepali, Khmer, Bangla).
• Type of Recording — Choose KII, FGD, or IDI. This affects the auto-generated prompt and the auto-selected AI model.
• AI Model — Automatically selected based on recording type and mode. Manual override is disabled.

Figure 3: Input dropdowns — Language, Recording Type, AI Model.

2

Review the Auto-Generated Prompt

Once you select a language and recording type, the User Prompt / Instructions text area auto-populates with a contextual prompt. This prompt tells the AI exactly how to transcribe — speaker labeling, timestamps, formatting rules, and quality notes.

You can edit this prompt to add context specific to your recording. For example:

• Number of participants if known
• Names of facilitators or interviewers
• Project name or topic area
• Any known audio quality issues

Figure 4: Auto-generated prompt for Nepali FGD.

💡

Pro Tip: Add participant names If you know the names of speakers, add them at the end of the prompt: "The interviewer is Priya and the respondent is Ramesh." The AI will use actual names instead of generic labels, making review much easier.

3

Upload Your Audio File

Drag and drop your audio file into the upload zone, or click to browse. The tool supports:

Format	Extension	Notes
MP3	.mp3	Most common format — recommended
WAV	.wav	Uncompressed, larger file size
M4A	.m4a	Apple/iPhone recordings
MP4	.mp4	Video files (audio track extracted)
WebM	.webm	Browser-based recordings
OGG / Opus	.ogg, .opus	Open-source formats
FLAC	.flac	Lossless audio

Maximum file size: 500 MB. Files over 100 MB will display a warning but will still process.

Figure 5: File preview after upload — validated and ready.

⚠

Large files > 100 MB For very large recordings, upload time depends on your internet speed. The progress bar will show real-time upload percentage. Keep the browser tab open until the upload completes.

4

Click "Transcribe"

Once all fields are valid, the Transcribe button activates. Click it to start processing. You will see:

1. A progress bar showing file upload percentage.
2. Status messages in the Process Log (dark area) — upload status, API calls, completion.
3. A spinner on the button while processing.

Processing time depends on file size and the effective model selected by policy. A 30-minute FGD recording typically takes 2–5 minutes with Gemini Pro.

Figure 6a: Upload progress bar and Processing button.

Figure 6b: Process Log — file uploaded, calling Gemini API.

5

Review & Download the Transcription

After processing completes, the results area appears with:

• Tab bar — Shows "📄 Original (Language)" tab.
• Result text area — The full transcription with speaker labels, timestamps, and notes.
• Action buttons:
  • Copy Text — Copies the transcription to your clipboard.
  • Download Docx File — Downloads the text as a .docx Word document.

Figure 7: Transcription result — Original (Nepali) with action buttons and Translate controls.

💡

Always download your transcription before translating Save the original transcription to your computer first. This gives you a backup in case anything goes wrong during translation, and you'll have both versions for your records.

6

Translate the Transcription (Optional)

After transcription, a translation control bar appears below the results. To translate:

1. Select your target language from the "Translate to:" dropdown.
2. Click Translate.
3. A new tab (e.g., "🌐 English") will appear next to the original tab.
4. Switch between tabs to compare original and translated text.

You can translate to multiple languages — each translation creates a new tab. Previously completed translations are cached so you don't need to re-process them.

Figure 8: Translation result — English tab active with translated text.

Doc Beta  Service 2 — Standalone Document Translation

Use this service when you already have a text document (report, transcript, field notes) and need it translated from one language to another — without any audio processing.

1

Switch to Document Translation Mode

Click the "Document Translation Beta" button in the mode switcher bar at the top. The form fields will change to show document-specific options.

Figure 9: Document Translation mode — source and target language selection.

3

Upload Your Document

The upload zone now accepts document formats instead of audio:

Format	Extension	Notes
Word Document	.docx	Best supported — formatting is partially preserved
PDF	.pdf	Text extracted; images/charts are not translated
Plain Text	.txt	Full text translated; no formatting to preserve

Figure 10: DOCX file uploaded and validated for translation.

4

Click "Translate Document"

The submit button now reads "Translate Document". Click it to start. For DOCX files, you'll see a multi-step process in the log:

1. Step 1: Reading and parsing the .docx file
2. Step 2: Translating the extracted text via AI
3. Step 3: Reconstructing the translated .docx file

A translated .docx file will automatically download when processing completes.

Figure 11: Document translation Process Log — parsing, extracting, translating.

AI Model Guide

The toolbox uses Gemini models only, with automatic model selection and policy-based overrides to balance quality, speed, and cost.

Supported Languages

The following languages are currently supported. This list may expand in the future.

Tips & Tricks for Best Results

🎤 Audio Recording Quality

• Record in a quiet environment when possible. Background noise (traffic, wind, market sounds) significantly reduces transcription accuracy.
• Use a directional microphone or place the recorder close to the speakers. Phone recordings held at arm's length often produce poor results.
• Brief participants to speak one at a time, especially in FGDs. Overlapping speech is the #1 cause of missed content.
• State participant names at the beginning of the recording (e.g., "This is the facilitator, Priya…"). This helps the AI label speakers correctly.

📝 Prompt Customization

• Add context: Mention the project name, topic area, and any relevant acronyms (e.g., "This is about the MMS project in Cambodia focusing on maternal nutrition").
• Specify speaker count: "There are 8 participants and 2 facilitators" helps the AI know how many voices to expect.
• Mention known issues: "The first 2 minutes have loud background music" or "Audio quality drops around the 30-minute mark" sets appropriate expectations.
• Don't ask for translation in the transcription prompt. Always transcribe first, review, then translate separately for best quality.

🔧 Model Selection Strategy

💾 Session Recovery

• The tool automatically saves your most recent transcription in your browser's local storage for up to 24 hours.
• If you accidentally close the tab or your browser crashes, reopening the tool will restore your last result automatically.
• Always download your transcription as a .docx to ensure you have a permanent copy.

🚀 Workflow Best Practices

1. Prepare your file first — ensure the audio format is supported and the file isn't corrupted before uploading.
2. Review the auto-prompt — take 30 seconds to read and optionally customize it. Small additions dramatically improve output.
3. Download before translating — save the original transcription first.
4. Compare translations side-by-side — use the tabs to switch between original and translated text for spot-checking.
5. For large batches, process one file at a time and save each result before starting the next.

Known Limitations

We believe in transparency. Here are the current limitations of the AI Toolbox:

Troubleshooting

Common Issues

Frequently Asked Questions

Q: Is my data stored anywhere by the AI providers?

Audio files are uploaded to Google's servers for processing and are subject to Google's data handling policies. Files are uploaded temporarily for processing and are not permanently stored by the tool itself. Check Helen Keller's data governance policy before processing sensitive or personally identifiable recordings.

Q: Can I use this on my phone or tablet?

The interface is responsive and works on mobile devices, but for the best experience (especially with large files), use your Helen Keller laptop or desktop with Chrome or Edge.

Q: Can I process a recording in a language not listed?

The Gemini AI models support many languages beyond the ones in the dropdown. You can select the closest available language and manually edit the prompt to specify the actual language. Results may vary for less common languages. If your country office regularly needs a language not listed, email AI@hki.org and we can add it to the dropdown.

Q: How do I report a bug or request a feature?

Email AI@hki.org or reach out to Ekendra or Ric directly. Include the browser you used, the file type, and any error messages from the Process Log — a screenshot of the log helps a lot.

Q: What happens if I close the browser mid-processing?

The processing will be interrupted and you will lose the in-progress result. The tool cannot resume interrupted sessions. You will need to re-upload and re-process the file.

Q: Can I edit the transcription before downloading?

The result text area is currently read-only. To edit, download the .docx file and edit in Microsoft Word or Google Docs, or copy the text and paste it into your preferred editor.

Q: Why is Document Translation marked as "Beta"?

The feature works reliably for text-based translation but has limited formatting preservation. We are actively improving it. Complex document layouts (tables, charts, multi-column) may not translate cleanly.

Quick Reference Card

A summary you can print or screenshot for quick reference: