An HR partner who speaks your language.

There's a second, gentler mode too: a text-and-voice team chat where the same Vanaja answers everyday questions and nudges people to build skills — sealing machines, garland-making, using the app — while staying inside the same confidentiality rails.

Claude Sonnet sits between the ears and the voice as the brain — it gets the salary context, the personality ("warm, caring, encouraging"), and the confidentiality rules, and produces the words. Sarvam makes those words sound like a person from down the road, not a call-centre robot.

The fix came from Sneha on the Sarvam team. Her advice was simple and exactly right: don't wait for the full answer — stream it. Speak each sentence the moment it's finished instead of holding the whole paragraph. Sarvam has a streaming TTS endpoint built for precisely this, and pairing it with Claude's token stream changes the feel of the whole thing.

// 1 · stream Claude; buffer tokens until a sentence ends
for await (const delta of claudeStream) {
  buffer += delta.text;
  const m = buffer.match(SENT_RE);          // . ! ? or । danda, ≥10 chars
  if (m) { flushSentence(m[1]); buffer = m[3]; }
}

// 2 · each finished sentence → Sarvam streaming TTS → mp3 chunks
async function _sarvamTtsStream(text, language, speaker, key) {
  const r = await fetch("https://api.sarvam.ai/text-to-speech/stream", {
    method: "POST",
    headers: { "api-subscription-key": key },
    body: JSON.stringify({
      text, target_language_code: language, speaker,
      model: "bulbul:v2", output_audio_codec: "mp3",
    }),
  });
  return { kind: "stream", body: r.body, mime: "audio/mpeg" };
}

// 3 · chain flushes so audio stays in order, pipe chunks to the browser as SSE
let pendingTail = Promise.resolve();
function flushSentence(sent) {
  pendingTail = pendingTail.then(() => flushSentenceImmediate(sent));
}

Time-to-first-word dropped from a multi-second wait to roughly 600ms — about an order of magnitude. That's the difference between a tool people tolerate and one they actually talk to. And it degrades gracefully: if streaming TTS isn't available it falls back to a single Sarvam WAV, and in the very last resort to the browser's own speech synthesis.

The shortest distance between a problem and the right person is often a public post.

The hard part of a vernacular voice agent isn't the languages — it's making it quick enough that someone forgets they're talking to software.