Skip to content

Speech-to-Text(STT)

Software that converts spoken audio — WhatsApp voice notes and calls — into text that bots, search and support agents can read and act on.

Also called: Voice note transcription · Automatic speech recognition (ASR) · Speech recognition

Reviewed

Interactive lab

Hands on: use the controls to explore. Anything that plays on its own pauses while you are working inside it.

Mask what a voice note said before you store it

Masking hides personal data before it reaches logs and dashboards, while keeping enough to recognise it.

Try

Text with email addresses or phone numbers.

Masked transcript
Haan ji, mera order number ••••8829 hai, delivery boy ko ••••• •3210 pe call karne bolo.
Emails
0
Numbers
2
  • Masked 2 phone or long numbers.
  • Numbers keep their last 4 digits — enough to confirm with a customer, useless to a fraudster.
  • Masking is one-way: do it before text reaches logs, analytics or exports.
Pick a raw transcript from a WhatsApp voice note and watch phone numbers, emails and long order numbers get masked before the text reaches logs or an LLM. Short values such as PIN codes stay readable.

01What is speech-to-text (STT)?

Speech-to-text (STT), also called automatic speech recognition (ASR), converts spoken audio into written text. For a WhatsApp business it is what lets a bot or an agent deal with a voice note as easily as a typed message: the customer records "mera order abhi tak nahi aaya", the system produces the transcript, and the rest of the pipeline — search, routing, replies — works on text.

WhatsApp does not transcribe for you on the API. A voice note arrives at your webhook as a message with type set to audio, whose audio object holds a media id, the mime_type (typically audio/ogg; codecs=opus) and a voice flag marking a recording made in WhatsApp. Your server exchanges the ID for a short-lived download URL, fetches the file with its access token and sends the audio to an STT engine. Calls through the WhatsApp Business Calling API can be transcribed too, but as streaming audio during the call rather than as a file afterwards.

STT is only the first step. It turns sound into words; natural language understanding (NLU) or a large language model then works out what those words mean. That split matters for accuracy: Indian customers switch between Hindi, English and regional languages mid-sentence, use local names and say numbers in different languages, so a transcript can read fluently and still be wrong in exactly the words that matter — an order number, a PIN code, a product name.

Key takeaways

  • Speech-to-text (STT) converts spoken audio into text, so a bot or agent can handle a voice note like a typed message.
  • On the Cloud API a voice note arrives as an audio message with a media ID; you download the file and transcribe it yourself.
  • Webhook media IDs expire after 7 days and download URLs after about 5 minutes, so fetch audio promptly.
  • Accuracy is measured as word error rate (WER) and drops with noise, strong accents and code-mixed Hindi-English speech.
  • STT only produces words; working out what they mean is the job of NLU or an LLM afterwards.

02 · How it works, step by step

How does speech-to-text (STT) work?

  1. Step 1

    Receive the voice note

    The messages webhook delivers a media message with type audio. Store the wamid, the sender's wa_id and the audio id, then acknowledge the webhook immediately — transcription runs in a background job.

  2. Step 2

    Download the audio

    Media retrieval takes two calls: GET /{MEDIA_ID} returns a url valid for about five minutes, and a GET on that URL with your bearer token returns the bytes. Media IDs received in webhooks expire after 7 days.

  3. Step 3

    Transcribe

    Send the OGG/Opus file, or a converted WAV, to your STT engine with language hints — or let it run language detection — and request word timestamps and confidence scores where the engine offers them.

  4. Step 4

    Normalise and mask the transcript

    Apply inverse text normalisation so spoken digits become numbers, then PII masking for phone numbers, emails and long account numbers before the text reaches logs, analytics or an LLM prompt.

  5. Step 5

    Hand the text on

    Pass the transcript to the bot as if the customer had typed it, show it to agents beside the audio player, and route low-confidence transcripts to a person or ask the customer to confirm key details.

03Speech-to-text (STT) in code

// Runs in a queue worker, after the webhook has been acknowledged.
async function transcribeVoiceNote(message) {
  const { id: mediaId, mime_type } = message.audio; // e.g. "audio/ogg; codecs=opus"
  const auth = { Authorization: `Bearer ${process.env.ACCESS_TOKEN}` };
 
  // 1. Exchange the media ID for a short-lived download URL (~5 minutes).
  const meta = await fetch(`https://graph.facebook.com/v26.0/${mediaId}`, { headers: auth })
    .then((r) => r.json());
 
  // 2. Download the bytes — the media URL needs the bearer token too.
  const audio = await fetch(meta.url, { headers: auth }).then((r) => r.arrayBuffer());
 
  // 3. Send to your speech-to-text engine (a generic HTTP API is shown).
  const form = new FormData();
  form.append("file", new Blob([audio], { type: mime_type }), "voice.ogg");
  form.append("language_hints", "hi,en");
  form.append("timestamps", "word");
  const stt = await fetch(process.env.STT_URL, { method: "POST", body: form })
    .then((r) => r.json()); // → { text, language, confidence, words: [...] }
 
  // 4. Mask personal data before the transcript is stored or sent to an LLM.
  return { wamid: message.id, from: message.from, ...stt, text: maskPii(stt.text) };
}

Two Graph API calls fetch the audio; the STT request and its fields depend on the engine you use.

04Example: Voice-note orders at a Chennai grocery

A neighbourhood grocery in Chennai takes orders on WhatsApp, and about half arrive as voice notes in a Tamil-English mix. Each note is downloaded within seconds of the webhook, transcribed with Tamil and English language hints, and masked before storage. The bot reads "two kilo tomato, one packet milk, address same as last time", builds a draft cart and replies with reply buttons to confirm. When the confidence on a product name is low, it asks "Did you mean *toor dal*?" instead of guessing, and voice notes it cannot parse at all go to the shop's agent with the audio attached.

05 · Common mistakes, and the fix

Common mistakes with speech-to-text (STT)

  • Transcribing inside the webhook request

    STT on a long voice note can take seconds. Doing it before returning 200 causes timeouts and retried, duplicated webhooks. Acknowledge first and transcribe from a queue.

  • Waiting too long to download

    The download URL lasts about five minutes and webhook media IDs expire after 7 days. Fetch audio as soon as the job runs, and request a fresh URL rather than reusing an old one.

  • Forcing a single language

    An English-only setting mangles Hinglish and regional speech; a Hindi-only setting mangles brand names. Use multi-language hints or automatic detection, and test on real customer audio.

  • Trusting numbers in transcripts

    Order IDs, amounts and phone numbers are where STT errors hurt most. Confirm critical values back to the customer, or validate them against your order system, before acting.

  • Storing raw transcripts with personal data

    Transcripts contain names, addresses and numbers spoken aloud. Mask them before logging, and apply the same retention rules to transcripts as to the audio under the DPDP Act.

06Speech-to-text (STT): key facts and limits

Inbound format
audio message, usually OGG with the Opus codec
Webhook media ID lifetime
7 days
Download URL lifetime
About 5 minutes
Cloud API audio limit
16 MB (AAC, AMR, MP3, M4A, OGG/Opus)
Accuracy metric
WER = (substitutions + deletions + insertions) ÷ words spoken
Voice notes vs calls
Batch file transcription vs streaming during the call

07Why does speech-to-text (STT) matter?

Voice notes are how many Indian customers prefer to talk — quicker than typing, easier in their own language, and natural for older or less text-comfortable users. A bot that replies "Sorry, I can only read text" loses them. With STT, voice notes enter the same queue, intent routing and search as text; agents can skim a transcript instead of replaying audio in a noisy office; and conversations become searchable for quality review and compliance.

08STT FAQs

Does the WhatsApp Cloud API transcribe voice notes automatically?+
No. The webhook carries the audio file's media ID, not a transcript; you download the file and transcribe it with a speech-to-text engine of your choice.
How accurate is speech-to-text for Indian languages?+
It varies widely by engine, language, audio quality and how much speakers mix languages. Measure word error rate on a sample of your own customers' voice notes; clean single-language audio scores far better than noisy, code-mixed speech.
What is the difference between speech-to-text and NLU?+
Speech-to-text turns audio into words; NLU works out what those words mean. A transcript reading "cancel kar do" is STT's output; recognising it as a cancellation request is NLU's job.
Can I transcribe WhatsApp calls as well as voice notes?+
Yes, with streaming speech-to-text on the call audio from the WhatsApp Business Calling API. Tell callers the call is transcribed, get consent where required, and handle transcripts as personal data.

09 · Tips from the bench

Best practices for speech-to-text (STT)

  • Keep the audio for a short, defined period so agents can check an unclear transcript.
  • Measure word error rate on a sample of your own voice notes per language, not on published benchmarks.
  • Show agents the transcript and the audio player together, with low-confidence words highlighted.
  • Reply in text even to voice notes unless you have tested text-to-speech in your customers' languages.

See InfiQ pricing

From definition to your first message

InfiQ is an official Meta Business Partner. Start a 7-day free trial — no card needed — and see this running on your own number.

Start free trialBrowse all 550 terms