search

willwade / tts-wrapper

TTS-Wrapper makes it easier to use text-to-speech APIs by providing a unified and easy-to-use interface.
MIT License
9 stars 2 forks source link

readme

py3-TTS-Wrapper

PyPI version codecov

Contributions are welcome! Check our contribution guide.

TTS-Wrapper simplifies using text-to-speech APIs by providing a unified interface across multiple services, allowing easy integration and manipulation of TTS capabilities.

Supported Services

Experimental (Not fully featured or in a state of WIP)

Features

Feature set overview

Engine OS Online/Offline SSML Rate/Volume/Pitch onWord events
Polly Linux/MacOS/Windows Online Yes Yes Yes
Google Linux/MacOS/Windows Online Yes Yes Yes
Azure Linux/MacOS/Windows Online Yes Yes Yes
Watson Linux/MacOS/Windows Online Yes No Yes
ElevenLabs Linux/MacOS/Windows Online No Yes Yes
Wit.AI Linux/MacOS/Windows Online Yes No No
Sherpa-Onnx Linux/MacOS/Windows Offline No No No
gTTS Linux/MacOS/Windows Online No No No
UWP Windows Offline No Yes No
SAPI Windows Offline Yes Yes Yes
NSS MacOS Offline Yes Yes Yes
eSpeak Linux/MacOS/Windows Offline No Yes No

Methods for each engine

Method Description Available Engines
speak() Plays synthesized speech directly. All engines
synth_to_file() Synthesizes speech and saves it to a file. All engines
speak_streamed() Streams synthesized speech. All engines
set_property() Sets properties like rate, volume, pitch. All engines
get_voices() Retrieves available voices. All engines
connect() Connects callback functions for events. Polly, Microsoft, Google, Watson.
pause_audio() Pauses ongoing speech playback. All engines
resume_audio() Resumes paused speech playback. All engines
stop_audio() Stops ongoing speech playback. All engines
set_output_device('id') Stops ongoing speech playback. All engines
check_credentials() True or False if Credentials are ok All engines

Notes:

Install

System Dependencies

This project requires the following system dependencies on Linux:

sudo apt-get insall portaudio19-dev

or MacOS, using Homebrew

brew install portaudio

For PicoTTS on Debian systems:

sudo apt-get install libttspico-utils

Using pip

pip install py3-tts-wrapper[google,microsoft,sapi,sherpaonnx,googletrans]

or via git

pip install git+https://github.com/willwade/tts-wrapper#egg=tts-wrapper[google,microsoft,sapi,mms,sherpaonnx]

or (the newer way we should all use)

pip install tts-wrapper[google,microsoft,sapi,sherpaonnx,googletrans]@git+https://github.com/willwade/tts-wrapper

NB: On MacOS(/zsh) you may need to do use quotes

pip install py3-tts-wrapper"[google, watson, polly, elevenlabs, microsoft, mms, sherpaonnx]"

Basic Usage

from tts_wrapper import PollyClient
pollyClient = PollyClient(credentials=('aws_key_id', 'aws_secret_access_key'))

from tts_wrapper import PollyTTS

tts = PollyTTS(pollyClient)
ssml_text = tts.ssml.add('Hello, <break time="500ms"/> world!')
tts.speak(ssml_text)

You can use SSML or plain text

from tts_wrapper import PollyClient
pollyClient = PollyClient(credentials=('aws_key_id', 'aws_secret_access_key'))
from tts_wrapper import PollyTTS

tts = PollyTTS(pollyClient)
tts.speak('Hello world')

For a full demo see the examples folder. You'll need to fill out the credentials.json (or credentials-private.json). Use them from cd'ing into the examples folder. Tips on gaining keys are below.

Authorization

Each service uses different methods for authentication:

Polly

from tts_wrapper import PollyTTS, PollyClient
client = PollyClient(credentials=('aws_region','aws_key_id', 'aws_secret_access_key'))

tts = PollyTTS(client)

Google

from tts_wrapper import GoogleTTS, GoogleClient
client = GoogleClient(credentials=('path/to/creds.json'))

tts = GoogleTTS(client)

or pass the auth file as dict - so in memory

from tts_wrapper import GoogleTTS, GoogleClient

with open(os.getenv("GOOGLE_CREDS_PATH"), "r") as file:
    credentials_dict = json.load(file)

client = GoogleClient(credentials=os.getenv('GOOGLE_CREDS_PATH'))
client = GoogleClient(credentials=credentials_dict)]

Microsoft

from tts_wrapper import MicrosoftTTS, MicrosoftClient
client = MicrosoftClient(credentials=('subscription_key','subscription_region'))

tts = MicrosoftTTS(client)

Watson

from tts_wrapper import WatsonTTS, WatsonClient
client = WatsonClient(credentials=('api_key', 'region', 'instance_id'))

tts = WatsonTTS(client)

Note If you have issues with SSL certification try

from tts_wrapper import WatsonTTS, WatsonClient
client = WatsonClient(credentials=('api_key', 'region', 'instance_id'),disableSSLVerification=True)

tts = WatsonTTS(client)

ElevenLabs

from tts_wrapper import ElevenLabsTTS, ElevenLabsClient
client = ElevenLabsClient(credentials=('api_key'))
tts = ElevenLabsTTS(client)

Wit.Ai

from tts_wrapper import WitAiTTS, WitAiClient
client = WitAiClient(credentials=('token'))
tts = WitAiTTS(client)

UWP

from tts_wrapper import UWPTTS, UWPClient
client = UWPClient()
tts = UWPTTS(client)

SAPI/eSpeak/NSSS

from tts_wrapper import SystemTTSClient, SystemTTS
client = SystemTTSClient('espeak') # eSpeak
client = SystemTTSClient('sapi') #SAPI
client = SystemTTSClient('nsss') #NSSS MacOS
# Initialize the TTS engine
tts = SystemTTSClient(client)

Just note: We cant do word timings in this.

GoogleTrans

Uses the gTTS library. 

from tts_wrapper import GoogleTransClient, GoogleTransTTS
voice_id = "en-co.uk"  # Example voice ID for UK English
client = GoogleTransClient(voice_id)
# Initialize the TTS engine
tts = GoogleTransTTS(client)

Sherpa-ONNX

You can provide blank model path and tokens path - and we will use a default location.. AS NOTED - WE HAVE DESIGNED THIS RIGHT NOW FOR MMS MODELS! We will add others like piper etc to this - Infact I'll drop regular piper support for sherpa-onnx. Its less of a headache..

from tts_wrapper import SherpaOnnxClient, SherpaOnnxTTS
client = SherpaOnnxClient(model_path=None, tokens_path=None)
tts = SherpaOnnxTTS(client)

Set a voice like

# Find voices/langs availables
voices = tts.get_voices()
print("Available voices:", voices)

# Set the voice using ISO code
iso_code = "eng"  # Example ISO code for the voice - also ID in voice details
tts.set_voice(iso_code)

and then use speak, speak_streamed etc..

You then can perform the following methods.

Advanced Usage

SSML

Even if you don't use SSML features that much its wise to use the same syntax - so pass SSML not text to all engines

ssml_text = tts.ssml.add('Hello world!')

Plain Text

If you want to keep things simple each engine will convert plain text to SSML if its not.

tts.speak('Hello World!')

Speak

This will use the default audio output of your device to play the audio immediately

tts.speak(ssml_text)

Check Credentials

This will check if the credentials are valid. Its only on the client object. Eg


    client = MicrosoftClient(
        credentials=(os.getenv("MICROSOFT_TOKEN"), os.getenv("MICROSOFT_REGION"))
    )
    if client.check_credentials():
        print("Credentials are valid.")
    else:
        print("Credentials are invalid."

NB: Each engine has a different way of checking credentials. If they dont have a supported the parent class will check get_voices. If you want to save calls just do a get_voices call.

Streaming and Playback Control

pause_audio(), resume_audio(), stop_audio()

These methods manage audio playback by pausing, resuming, or stopping it. NB: Only to be used for speak_streamed

You need to make sure the optional dependency is included for this

pip install py3-tts-wrapper[controlaudio,google.. etc

then


client = GoogleClient(..)
tts = GoogleTTS(client)
try:
    text = "This is a pause and resume test. The text will be longer, depending on where the pause and resume works"
    audio_bytes = tts.synth_to_bytes(text)
    tts.load_audio(audio_bytes)
    print("Play audio for 3 seconds")
    tts.play(1)
    tts.pause(8)
    tts.resume()
    time.sleep(6)
finally:
    tts.cleanup()

NB: to do this we use pyaudio. If you have issues with this you may need to install portaudio19-dev - particularly on linux

sudo apt-get install portaudio19-dev

File Output

tts.synth_to_file(ssml_text, 'output.mp3', format='mp3')

there is also "synth" method which is legacy. Note we support saving as mp3, wav or flac. 

tts.synth('<speak>Hello, world!</speak>', 'hello.mp3', format='mp3)

Note you can also stream - and save. Just note it saves at the end of streaming entirely..

ssml_text = tts.ssml.add('Hello world!')

tts.speak_streamed(ssml_text,filepath,'wav')

Fetch Available Voices

voices = tts.get_voices()
print(voices)

NB: All voices will have a id, dict of language_codes, name and gender. Just note not all voice engines provide gender

Voice Selection

tts.set_voice(voice_id,lang_code=en-US)

e.g.

tts.set_voice('en-US-JessaNeural','en-US')

Use the id - not a name

SSML

ssml_text = tts.ssml.add('Hello, <break time="500ms"/> world!')
tts.speak(ssml_text)

Volume, Rate and Pitch Control

Set volume:

tts.set_property("volume", "90")
text_read = f"The current volume is 90"
text_with_prosody = tts.construct_prosody_tag(text_read)
ssml_text = tts.ssml.add(text_with_prosody)

Set rate:

tts.set_property("rate", "slow")
text_read = f"The current rate is SLOW"
text_with_prosody = tts.construct_prosody_tag(text_read)
ssml_text = tts.ssml.add(text_with_prosody)

Speech Rate:

Set pitch:

tts.set_property("pitch", "high")
text_read = f"The current pitch is SLOW"
text_with_prosody = tts.construct_prosody_tag(text_read)
ssml_text = tts.ssml.add(text_with_prosody)

Pitch Control:

Use the tts.ssml.clear_ssml() method to clear all entries from the ssml list

set_property()

This method allows setting properties like rate, volume, and pitch.

tts.set_property("rate", "fast")
tts.set_property("volume", "80")
tts.set_property("pitch", "high")

get_property()

This method retrieves the value of properties such as volume, rate, or pitch.

current_volume = tts.get_property("volume")
print(f"Current volume: {current_volume}")

Using callbacks on word-level boundaries

Note only Polly, Microsoft, Google, ElevenLabs, UWP, SAPI and Watson can do this correctly. We can't do this in anything else but we do do a estimated tonings for all other engines (ie elevenlabs, witAi and Piper)

def my_callback(word: str, start_time: float, end_time: float):
    duration = end_time - start_time
    print(f"Word: {word}, Duration: {duration:.3f}s")

def on_start():
    print('Speech started')

def on_end():
    print('Speech ended')

try:
    text = "Hello, This is a word timing test"
    ssml_text = tts.ssml.add(text)
    tts.connect('onStart', on_start)
    tts.connect('onEnd', on_end)
    tts.start_playback_with_callbacks(ssml_text, callback=my_callback)
except Exception as e:
    print(f"Error: {e}")

and it will output

Speech started
Word: Hello, Duration: 0.612s
Word: , Duration: 0.212s
Word: This, Duration: 0.364s
Word: is, Duration: 0.310s
Word: a, Duration: 0.304s
Word: word, Duration: 0.412s
Word: timing, Duration: 0.396s
Word: test, Duration: 0.424s
Speech ended

connect()

This method allows registering callback functions for events like onStart or onEnd.

def on_start():
    print("Speech started")

tts.connect('onStart', on_start)

Supported File Formats

By default, all engines output audio in the WAV format, but can be configured to output MP3 or other formats where supported.

tts.synth('<speak>Hello, world!</speak>', 'hello.mp3', format='mp3)

The synth_to_bytestream method is designed to synthesize text into an in-memory bytestream in the specified audio format (wav, mp3, flac, etc.). It is particularly useful when you want to handle the audio data in-memory for tasks like saving it to a file, streaming the audio, or passing it to another system for processing.

Method Signature:

def synth_to_bytestream(self, text: Any, format: Optional[str] = "wav") -> BytesIO:
    """
    Synthesizes text to an in-memory bytestream in the specified audio format.

    :param text: The text to synthesize.
    :param format: The audio format (e.g., 'wav', 'mp3', 'flac'). Default: 'wav'.
    :return: A BytesIO object containing the audio data.
    """

Parameters:

Returns:

Example Use Cases

1. Saving Audio to a File

You can use the synth_to_bytestream method to synthesize audio in any supported format and save it directly to a file.

# Synthesize text into a bytestream in MP3 format
bytestream = tts.synth_to_bytestream("Hello, this is a test", format="mp3")

# Save the audio bytestream to a file
with open("output.mp3", "wb") as f:
    f.write(bytestream.read())

print("Audio saved to output.mp3")

Explanation:

2. Real-Time Playback Using sounddevice

If you want to play the synthesized audio live without saving it to a file, you can use the sounddevice library to directly play the audio from the BytesIO bytestream.

import sounddevice as sd
import numpy as np

# Synthesize text into a bytestream in WAV format
bytestream = tts.synth_to_bytestream("Hello, this is a live playback test", format="wav")

# Convert the bytestream back to raw PCM audio data for playback
audio_data = np.frombuffer(bytestream.read(), dtype=np.int16)

# Play the audio using sounddevice
sd.play(audio_data, samplerate=tts.audio_rate)
sd.wait()

print("Live playback completed")

Explanation:

Developer's Guide

Setting up the Development Environment

Using Pipenv

  1. Clone the repository:

    git clone https://github.com/willwade/tts-wrapper.git
cd tts-wrapper

  2. Install the package and system dependencies:

    pip install .

    To install optional dependencies, use:

    pip install .[google, watson, polly, elevenlabs, microsoft]

This will install Python dependencies and system dependencies required for this project. Note that system dependencies will only be installed automatically on Linux.

Using Poetry

  1. Clone the repository:

    git clone https://github.com/willwade/tts-wrapper.git
cd tts-wrapper

  2. Install Python dependencies:

    poetry install

  3. Install system dependencies (Linux only):

    poetry run postinstall

NOTE: to get a requirements.txt file for the project use poetry export --without-hashes --format=requirements.txt > requirements.txt --all-extras juat be warned that this will include all dependencies including dev ones.

Release a new build

git tag -a v0.1.0 -m "Release 0.1.0"
git push origin v0.1.0

Adding a New Engine to TTS Wrapper

This guide provides a step-by-step approach to adding a new engine to the existing Text-to-Speech (TTS) wrapper system.

Step 1: Create Engine Directory Structure

  1. Create a new folder for your engine within the engines directory. Name this folder according to your engine, such as witai for Wit.ai.

    Directory structure:

    engines/witai/

  2. Create necessary files within this new folder:

    • __init__.py - Makes the directory a Python package.
    • client.py - Handles all interactions with the TTS API.
    • engine.py - Contains the TTS class that integrates with your abstract TTS system.
    • ssml.py - Defines any SSML handling specific to this engine.

    Final directory setup:

    engines/
└── witai/
   ├── __init__.py
   ├── client.py
   ├── engine.py
   └── ssml.py

Step 2: Implement Client Functionality in client.py

Implement authentication and necessary setup for API connection. This file should manage tasks such as sending synthesis requests and fetching available voices.

class TTSClient:
    def __init__(self, api_key):
        self.api_key = api_key
        # Setup other necessary API connection details here

    def synth(self, text, options):
        # Code to send a synthesis request to the TTS API
        pass

    def get_voices(self):
        # Code to retrieve available voices from the TTS API
        pass

Step 3: Define the TTS Engine in engine.py

This class should inherit from the abstract TTS class and implement required methods such as get_voices and synth_to_bytes.

from .client import TTSClient
from your_tts_module.abstract_tts import AbstractTTS

class WitTTS(AbstractTTS):
    def __init__(self, api_key):
        super().__init__()
        self.client = TTSClient(api_key)

    def get_voices(self):
        return self.client.get_voices()

    def synth_to_bytes(self, text, format='wav'):
        return self.client.synth(text, {'format': format})

Step 4: Implement SSML Handling in ssml.py

If the engine has specific SSML requirements or supports certain SSML tags differently, implement this logic here.

from your_tts_module.abstract_ssml import BaseSSMLRoot, SSMLNode

class EngineSSML(BaseSSMLRoot):
    def add_break(self, time='500ms'):
        self.root.add(SSMLNode('break', attrs={'time': time}))

Step 5: Update __init__.py

Make sure the __init__.py file properly imports and exposes the TTS class and any other public classes or functions from your engine.

from .engine import WitTTS
from .ssml import EngineSSML

Tips

Getting keys

Watson

This is not straightforward

Polly

Microsoft

Google

Create a Service Account:

  1. Go to the Google Cloud Console: Visit the Google Cloud Console.
  2. Create a New Project: If you don't already have a project, create a new one in the developer console.
  3. Enable APIs: Enable the APIs that your service account will be using. For example, if you're using Google Drive API, enable that API for your project.
  4. Create a Service Account:
  1. Create and Download Credentials:

Wit.Ai

  1. https://wit.ai/apps
  2. Look for Bearer token. Its in the Curl example

ElevenLabs

  1. Login at https://elevenlabs.io/app/speech-synthesis
  2. Go to your profile and click on "Profile + API Key"
  3. Click on Popup and copy "API Key"

License

This project is licensed under the MIT License.