Made in Vancouver, Canada by Picovoice

Orca is an on-device streaming text-to-speech engine that is designed for use with LLMs, enabling zero-latency voice assistants. Orca is:

• Private; All voice processing runs locally.
• Cross-Platform:
  • Linux (x86_64), macOS (x86_64, arm64), Windows (x86_64)
  • Android and iOS
  • Raspberry Pi (5, 4, 3) and NVIDIA Jetson Nano
  • Chrome, Safari, Firefox, and Edge

Please note that Orca is currently in development. While we prioritize stability and compatibility, certain aspects of Orca may undergo changes as we continually enhance and refine the engine to provide the best user experience possible.

• Orca
  • Table of Contents
  • Overview
    • Orca streaming text synthesis
    • Text input
    • Custom pronunciations
    • Voices
    • Speech control
    • Audio output
  • AccessKey
  • Demos
    • Python Demos
    • iOS Demo
    • C Demos
    • Web Demos
    • Android Demo
  • SDKs
    • Python
    • iOS
    • C
    • Web
    • Android
  • Releases
  • FAQ

• Orca Text-to-Speech currently supports English only.
• Support for additional languages is available for commercial customers on a case-by-case basis.

Orca is a text-to-speech engine designed specifically for LLMs. It can process incoming text streams in real-time, generating audio continuously, i.e., as the LLM produces tokens, Orca generates speech in parallel. This enables seamless conversations with voice assistants, eliminating any audio delays.

Orca also supports single synthesis mode, where a complete text is synthesized in a single call to the Orca engine.

Orca accepts the 26 lowercase (a-z) and 26 uppercase (A-Z) letters of the English alphabet, numbers, basic symbols, as well as common punctuation marks. You can get a list of all supported characters by calling the valid_characters() method provided in the Orca SDK you are using. Pronunciations of characters or words not supported by this list can be achieved with custom pronunciations.

Orca supports custom pronunciations via a specific syntax embedded within the input text. This feature allows users to define unique pronunciations for words using the following format: {word|pronunciation}. The pronunciation is expressed in ARPAbet phonemes. The following are examples of sentences using custom pronunciations:

• "This is a {custom|K AH S T AH M} pronunciation"
• "{read|R IY D} this as {read|R EH D}, please."
• "I {live|L IH V} in {Sevilla|S EH V IY Y AH}. We have great {live|L AY V} sports!"

Orca can synthesize speech with various voices, each of which is characterized by a model file located in lib/common. To synthesize speech with a specific voice, provide the associated model file as an argument to the orca init function. The following are the voices currently available:

Orca provides a set of parameters to control the synthesized speech. The following table lists the available parameters:

Orca's synthesized speech is delivered as either raw audio data or a WAV file. Output audio will be in single-channel 16-bit PCM format and can be directly fed into a playback audio system.

AccessKey is your authentication and authorization token for deploying Picovoice SDKs, including Orca. Anyone who is using Picovoice needs to have a valid AccessKey. You must keep your AccessKey secret. You will need internet connectivity to validate your AccessKey with Picovoice license servers even though the text-to-speech engine is running 100% offline.

AccessKey also verifies that your usage is within the limits of your account. Everyone who signs up for Picovoice Console receives the Free Tier usage rights described here. If you wish to increase your limits, you can purchase a subscription plan.

Install the demo package:

pip3 install pvorcademo

Run the streaming demo:

orca_demo_streaming --access_key ${ACCESS_KEY} --text_to_stream ${TEXT}

Run the single synthesis demo:

orca_demo --access_key ${ACCESS_KEY} --text ${TEXT} --output_path ${WAV_OUTPUT_PATH}

Replace ${ACCESS_KEY} with yours obtained from Picovoice Console, ${TEXT} with the text to be synthesized, and ${WAV_OUTPUT_PATH} with a path to an output WAV file.

For more information about Python demos go to demo/python.

Run the following from demo/ios to install the Orca-iOS CocoaPod:

pod install

Replace let ACCESS_KEY = "..." inside ViewModel.swift with yours obtained from Picovoice Console.

Then, using Xcode, open the generated OrcaDemo.xcworkspace and run the application.

For more information about iOS demos go to demo/ios.

Build the streaming demo:

cmake -S demo/c/ -B demo/c/build && cmake --build demo/c/build --target orca_demo_streaming

Run the demo:

./demo/c/build/orca_demo_streaming -l ${LIBRARY_PATH} -m ${MODEL_PATH} -a ${ACCESS_KEY} -t ${TEXT} -o ${OUTPUT_PATH}

From demo/web run the following in the terminal:

yarn
yarn start

(or)

npm install
npm run start

Open http://localhost:5000 in your browser to try the demo.

Using Android Studio, open demo/android/OrcaDemo as an Android project and then run the application.

Replace "${YOUR_ACCESS_KEY_HERE}" in the file MainActivity.java with your AccessKey.

Install the Python SDK:

pip3 install pvorca

Create an instance of the engine:

import pvorca

orca = pvorca.create(access_key='${ACCESS_KEY}')

Replace ${ACCESS_KEY} with yours obtained from Picovoice Console.

To synthesize a text stream, create an Orca Stream object and add text to it one-by-one:

stream = orca.stream_open()

for text_chunk in text_generator():
    pcm = stream.synthesize(text_chunk)
    if pcm is not None:
        # handle pcm

pcm = stream.flush()
if pcm is not None:
    # handle pcm

The text_generator() function can be any stream generating text, for example an LLM response. When done with streaming text synthesis, the stream object needs to be closed:

stream.close()

Use single synthesis mode if the complete text is known in advance:

pcm, alignments = orca.synthesize('${TEXT}')

Replace ${TEXT} with the text to be synthesized including potential custom pronunciations.

Finally, when done make sure to explicitly release the resources:

orca.delete()

For more details see Python SDK.

Create an instance of the engine:

import Orca

let modelPath = Bundle(for: type(of: self)).path(
        forResource: "${MODEL_FILE}", // Name of the model file name for Orca
        ofType: "pv")!

do {
  let orca = try Orca(accessKey: "${ACCESS_KEY}", modelPath: modelPath)
} catch {}

Replace ${ACCESS_KEY} with yours obtained from Picovoice Console and ${MODEL_FILE} with the model file name for Orca.

To synthesize a text stream, create an OrcaStream object and add text to it one-by-one via the synthesize method. Call flush to synthesize any remaining text, and close to delete the OrcaStream object:

let orcaStream = try orca.streamOpen()

for textChunk in textGenerator() {
  let pcm = orcaStream.synthesize(textChunk)
  if pcm != nil {
    // handle pcm
  }
}

let pcm = orcaStream.flush()
if pcm != nil {
  // handle pcm
}

orcaStream.close()

textGenerator() can be any stream generating text, for example an LLM response.

do {
    let (pcm, wordArray) = try orca.synthesize(text: "${TEXT}")
} catch {}

Replace ${TEXT} with the text to be synthesized including potential custom pronunciations.

When done be sure to explicitly release the resources using orca.delete().

For more details, see the iOS SDK.

The header file include/pv_orca.h contains relevant information on Orca's C SDK.

Build an instance of the object:

pv_orca_t *orca = NULL;
const char *model_path = "${MODEL_PATH}";
pv_status_t status = pv_orca_init("${ACCESS_KEY}", model_path, &orca);
if (status != PV_STATUS_SUCCESS) {
    // error handling logic
}

Replace ${ACCESS_KEY} with the AccessKey obtained from Picovoice Console, and ${MODEL_PATH} with the path to the model file available under lib/common.

Create a synthesize_params object to control the synthesized speech:

pv_orca_synthesize_params_t *synthesize_params = NULL;
status = pv_orca_synthesize_params_init(&synthesize_params);
// change the default parameters of synthesize_params as desired

To synthesize a text stream, create an orca_stream object using synthesize_params:

pv_orca_stream_t *orca_stream = NULL;
status = pv_orca_stream_open(orca, synthesize_params, &orca_stream);
if (status != PV_STATUS_SUCCESS) {
    // error handling logic
}

Add text to orca_stream one-by-one and handle the synthesized audio:

extern char *get_next_text_chunk(void);

int32_t num_samples_chunk = 0;
int16_t *pcm_chunk = NULL;
status = pv_orca_stream_synthesize(
    orca_stream, 
    get_next_text_chunk(), 
    &num_samples_chunk, 
    &pcm_chunk);
if (status != PV_STATUS_SUCCESS) {
    // error handling logic
}
if (num_samples_chunk > 0) {
    // handle pcm_chunk
}

Once the text stream is complete, call the flush method to synthesize the remaining text:

status = pv_orca_stream_flush(orca_stream, &num_samples_chunk, &pcm_chunk);
if (status != PV_STATUS_SUCCESS) {
    // error handling logic
}
if (num_samples_chunk > 0) {
    // handle pcm_chunk
}

Once the PCM chunks are handled, make sure to release the acquired resources for each chunk with:

pv_orca_pcm_delete(pcm_chunk);

Finally, when done make sure to close the stream:

pv_orca_stream_close(orca_stream);

If the text is known in advance, single synthesis mode can be used:

int32_t num_samples = 0;
int16_t *synthesized_pcm = NULL;
int32_t num_alignments = 0;
pv_orca_word_alignment_t **alignments = NULL;
status = pv_orca_synthesize(
    orca,
    "${TEXT}",
    synthesize_params,
    &num_samples,
    &synthesized_pcm,
    &num_alignments,
    &alignments);

Replace ${TEXT} with the text to be synthesized including potential custom pronunciations.

Print the metadata of the synthesized audio:

for (int32_t i = 0; i < num_alignments; i++) {
    fprintf(
            stdout,
            "[%s]\t.start_sec = %.2f .end_sec = %.2f\n",
            alignments[i].word,
            alignments[i].start_sec,
            alignments[i].end_sec);
    for (int32_t j = 0; j < alignments[i].num_phonemes; j++) {
        fprintf(
                stdout,
                "\t[%s]\t.start_sec = %.2f .end_sec = %.2f\n",
                alignments[i].phonemes[j].phoneme,
                alignments[i].phonemes[j].start_sec,
                alignments[i].phonemes[j].end_sec);
    
    }
}

Finally, when done make sure to release the acquired resources:

pv_orca_word_alignments_delete(num_alignments, alignments);
pv_orca_pcm_delete(pcm);
pv_orca_synthesize_params_delete(synthesize_params);
pv_orca_delete(orca);

Install the web SDK using yarn:

yarn add @picovoice/orca-web

or using npm:

npm install --save @picovoice/orca-web

Create an instance of the engine using OrcaWorker and synthesize speech:

import { OrcaWorker } from "@picovoice/orca-web";
import orcaParams from "${PATH_TO_BASE64_ORCA_PARAMS}";

const orca = await OrcaWorker.create(
  "${ACCESS_KEY}",
  { base64: orcaParams }
);

Replace ${ACCESS_KEY} with yours obtained from Picovoice Console.

To synthesize a text stream, create an OrcaStream object and add text to it one-by-one via the synthesize method. Call flush to synthesize any remaining text, and close to delete the OrcaStream object:

const orcaStream = await orca.streamOpen();

function* textStream(): IterableIterator<string> {
  ... // yield text chunks e.g. from an LLM response
}

for (const textChunk of textStream()) {
  const pcm = await orcaStream.synthesize(textChunk);
  if (pcm !== null) {
    // handle pcm
  }
}

const flushedPcm = orcaStream.flush();
if (flushedPcm !== null) {
  // handle pcm
}

orcaStream.close();

const { speechPcm, alignments } = await orca.synthesize("${TEXT}")

Finally, when done release the resources using orca.release().

For more details, see the Web SDK.

To include the Orca package in your Android project, ensure you have included mavenCentral() in your top-level build.gradle file and then add the following to your app's build.gradle:

dependencies {
    implementation 'ai.picovoice:orca-android:${LATEST_VERSION}'
}

Create an instance of the engine and generate speech:

import ai.picovoice.orca.*;

final String accessKey = "${ACCESS_KEY}"; // AccessKey obtained from Picovoice Console (https://console.picovoice.ai/)
final String modelPath = "${MODEL_FILE_PATH}";
try {
    Orca orca = new Orca.Builder()
        .setAccessKey(accessKey)
        .setModelPath(modelPath)
        .build(appContext);
} catch (OrcaException ex) { }

Replace ${ACCESS_KEY} with yours obtained from Picovoice Console, ${MODEL_FILE_PATH} with an Orca voice model file and ${TEXT} with the text to be synthesized including potential custom pronunciations.

To synthesize a text stream, create an OrcaStream object and add text to it one-by-one via the synthesize. Call flush to synthesize any remaining text, and close to delete the OrcaStream object:

Orca.OrcaStream orcaStream = orca.streamOpen(new OrcaSynthesizeParams.Builder().build());

for (String textChunk : textGenerator()) {
  short[] pcm = orcaStream.synthesize(textChunk);
  if (pcm != null) {
    // handle pcm
  }
}

short[] flushedPcm = orcaStream.flush();
if (flushedPcm != null) {
  // handle pcm
}

OrcaAudio audio = orca.synthesize(
    "${TEXT}",
    new OrcaSynthesizeParams.Builder().build());

Finally, when done make sure to explicitly release the resources:

orca.delete()

For more details, see the Android SDK.

• Support for streaming synthesis
• Reduced model size
• Improved performance
• Support for word alignments
• Improved naturalness and pronunciations

• Beta release

You can find the FAQ here.