Google Cloud Speech Free code examples in C#, Python, Java and JavaScript

No API key. Google Cloud uses Application Default Credentials — either a service-account file, or `gcloud auth application-default login` on a laptop.

pip install google-cloud-texttospeech

# A service account, for a server:
export GOOGLE_APPLICATION_CREDENTIALS="/path/to/service-account.json"

# Or, on a developer machine:
gcloud auth application-default login

The locale goes in TWO places and they must agree: language_code on the voice, and the locale prefix of the name. A mismatch is accepted and produces an accent nobody asked for.

from google.cloud import texttospeech

client = texttospeech.TextToSpeechClient()

response = client.synthesize_speech(
    input=texttospeech.SynthesisInput(
        text="नमस्ते, यह हिंदी में एक परीक्षण है।"
    ),
    voice=texttospeech.VoiceSelectionParams(
        language_code="hi-IN",
        name="hi-IN-Chirp3-HD-Charon",
    ),
    audio_config=texttospeech.AudioConfig(
        audio_encoding=texttospeech.AudioEncoding.MP3
    ),
)

with open("hindi.mp3", "wb") as handle:
    handle.write(response.audio_content)

DO NOT HARDCODE A VOICE LIST. Availability changes per language and per region, and list_voices is the only answer that is true on the day you run it — the same reasoning as the Bedrock recipe's model discovery step.

voices = client.list_voices(language_code="hi-IN")

for voice in voices.voices:
    # Chirp 3 HD voices are the ones with Chirp3-HD in the name; the
    # rest are Standard, WaveNet, Neural2 or Studio, which have
    # different capabilities and different pricing.
    if "Chirp3-HD" in voice.name:
        print(voice.name, voice.ssml_gender.name, voice.natural_sample_rate_hertz)

Charon is Charon in every locale that has it. Swapping the prefix moves language without changing how the voice sounds, which is what keeps a multilingual product from sounding like four different products.

GREETINGS = {
    "hi-IN": "नमस्ते, आपका स्वागत है।",
    "en-IN": "Hello, and welcome.",
    "bn-IN": "নমস্কার, স্বাগতম।",
    "ta-IN": "வணக்கம், வரவேற்கிறோம்.",
}

for locale, text in GREETINGS.items():
    response = client.synthesize_speech(
        input=texttospeech.SynthesisInput(text=text),
        voice=texttospeech.VoiceSelectionParams(
            language_code=locale,
            name=f"{locale}-Chirp3-HD-Charon",
        ),
        audio_config=texttospeech.AudioConfig(
            audio_encoding=texttospeech.AudioEncoding.MP3
        ),
    )

    with open(f"greeting-{locale}.mp3", "wb") as handle:
        handle.write(response.audio_content)

The BOM keeps the client and its gRPC transitive dependencies on versions that agree with each other — skipping it is the usual cause of a NoSuchMethodError at startup.

# Gradle
implementation(platform("com.google.cloud:libraries-bom:26.66.0"))
implementation("com.google.cloud:google-cloud-texttospeech")

export GOOGLE_APPLICATION_CREDENTIALS="/path/to/service-account.json"

The client is AutoCloseable and expensive to build — one per process, not one per request. ByteString gives you the bytes without another copy.

import com.google.cloud.texttospeech.v1.*;
import com.google.protobuf.ByteString;

import java.io.FileOutputStream;
import java.io.OutputStream;

public class SpeakHindi {
    public static void main(String[] args) throws Exception {
        try (TextToSpeechClient client = TextToSpeechClient.create()) {
            SynthesisInput input = SynthesisInput.newBuilder()
                .setText("नमस्ते, यह हिंदी में एक परीक्षण है।")
                .build();

            VoiceSelectionParams voice = VoiceSelectionParams.newBuilder()
                .setLanguageCode("hi-IN")
                .setName("hi-IN-Chirp3-HD-Charon")
                .build();

            AudioConfig audio = AudioConfig.newBuilder()
                .setAudioEncoding(AudioEncoding.MP3)
                .build();

            SynthesizeSpeechResponse response =
                client.synthesizeSpeech(input, voice, audio);

            ByteString content = response.getAudioContent();
            try (OutputStream out = new FileOutputStream("hindi.mp3")) {
                content.writeTo(out);
            }
        }
    }
}

Which Chirp 3 HD voices exist for a given Indic locale changes. Ask, rather than shipping a constant that was accurate once.

try (TextToSpeechClient client = TextToSpeechClient.create()) {
    ListVoicesResponse voices = client.listVoices("hi-IN");

    for (Voice voice : voices.getVoicesList()) {
        if (!voice.getName().contains("Chirp3-HD")) {
            continue;
        }

        System.out.printf("%s  %s  %dHz%n",
            voice.getName(),
            voice.getSsmlGender(),
            voice.getNaturalSampleRateHertz());
    }
}

Same character, four languages. This is the whole trick to sounding like one product in a country with twenty-two official languages.

import java.util.Map;

Map<String, String> greetings = Map.of(
    "hi-IN", "नमस्ते, आपका स्वागत है।",
    "en-IN", "Hello, and welcome.",
    "bn-IN", "নমস্কার, স্বাগতম।",
    "ta-IN", "வணக்கம், வரவேற்கிறோம்.");

try (TextToSpeechClient client = TextToSpeechClient.create()) {
    for (Map.Entry<String, String> entry : greetings.entrySet()) {
        String locale = entry.getKey();

        SynthesizeSpeechResponse response = client.synthesizeSpeech(
            SynthesisInput.newBuilder().setText(entry.getValue()).build(),
            VoiceSelectionParams.newBuilder()
                .setLanguageCode(locale)
                .setName(locale + "-Chirp3-HD-Charon")
                .build(),
            AudioConfig.newBuilder().setAudioEncoding(AudioEncoding.MP3).build());

        try (OutputStream out = new FileOutputStream("greeting-" + locale + ".mp3")) {
            response.getAudioContent().writeTo(out);
        }
    }
}

The npm package is scoped. Credentials come from the environment, so nothing about the key lives in code.

npm install @google-cloud/text-to-speech

export GOOGLE_APPLICATION_CREDENTIALS="/path/to/service-account.json"

synthesizeSpeech returns an array — the response is the first element. Destructuring it, as below, is the idiom in every Google Cloud client and it catches people out once.

import { writeFile } from "node:fs/promises";
import textToSpeech from "@google-cloud/text-to-speech";

const client = new textToSpeech.TextToSpeechClient();

const [response] = await client.synthesizeSpeech({
  input: { text: "नमस्ते, यह हिंदी में एक परीक्षण है।" },
  voice: {
    languageCode: "hi-IN",
    name: "hi-IN-Chirp3-HD-Charon",
  },
  audioConfig: { audioEncoding: "MP3" },
});

await writeFile("hindi.mp3", response.audioContent, "binary");

A hardcoded voice list goes stale the first time Google adds a language. listVoices is one call and it is correct by construction.

const [available] = await client.listVoices({ languageCode: "hi-IN" });

for (const voice of available.voices) {
  if (!voice.name.includes("Chirp3-HD")) continue;

  console.log(voice.name, voice.ssmlGender, voice.naturalSampleRateHertz);
}

speakingRate, pitch and SSML input are accepted and have NO EFFECT on a Chirp 3 HD voice. No error is raised. If you need them, that is a reason to pick a Neural2 or Studio voice instead.

// This runs, returns audio, and ignores both knobs.
const [ignored] = await client.synthesizeSpeech({
  input: { text: "This is read at the default rate and pitch." },
  voice: { languageCode: "hi-IN", name: "hi-IN-Chirp3-HD-Charon" },
  audioConfig: {
    audioEncoding: "MP3",
    speakingRate: 1.4,   // ignored by Chirp 3 HD
    pitch: -4.0,         // ignored by Chirp 3 HD
  },
});

// A Neural2 voice honours both, at the cost of a less natural read.
const [controlled] = await client.synthesizeSpeech({
  input: { text: "This one is genuinely faster." },
  voice: { languageCode: "hi-IN", name: "hi-IN-Neural2-A" },
  audioConfig: { audioEncoding: "MP3", speakingRate: 1.4 },
});