2024.03.08

Building a Nepali Text-to-Speech API.

Wrapping Google Cloud TTS and a custom Nepali voice model into a simple REST API for production use.

TTSAPINepalNode.js

Nepali TTS is a solved problem in theory — Google Cloud TTS has a standard Nepali voice — but getting it into a production API that handles Unicode correctly and doesn't break on mixed script input takes some work.

Setup

npm init -y
npm install express @google-cloud/text-to-speech dotenv
# Set GOOGLE_APPLICATION_CREDENTIALS in .env

Core TTS endpoint

const textToSpeech = require('@google-cloud/text-to-speech');
const client = new textToSpeech.TextToSpeechClient();

app.post('/synthesize', async (req, res) => {
  const { text } = req.body;
  const [response] = await client.synthesizeSpeech({
    input: { text },
    voice: { languageCode: 'ne-NP', ssmlGender: 'FEMALE' },
    audioConfig: { audioEncoding: 'MP3' },
  });
  res.set('Content-Type', 'audio/mpeg');
  res.send(response.audioContent);
});

Gotchas with Nepali Unicode

Devanagari text must be NFC-normalized before synthesis
Mixed Nepali-English input confuses the language model — split and synthesize separately
Some conjunct characters render incorrectly in older Google voice models
The 'ne-NP' voice doesn't support SSML — keep input as plain text

What this API powers

→Offline pronunciation guides for Nepali vocabulary apps
→Voice notifications for IoT sensor alerts in rural areas
→Read-aloud mode for a Nepali news aggregator
→Accessibility layer for forms in government portals

For higher-quality output, OpenAI TTS can be used with Romanized Nepali — the transcription quality is better, though it loses the native accent. For a native voice, Google remains the practical choice as of 2024.