This repository contains a fully featured sample script to demonstrate how to connect to the Resemble.AI Live STS server using Socket.IO for real-time speech-to-speech.

• Installation
• Usage
• Arguments
• Developing Custom Socket Client
• License

1. Clone the repository:

   git clone https://github.com/resemble-ai/resemble-live-sts-socket.git  
   cd resemble-live-sts-socket

2. Install the required dependencies:

   conda create -n socket_demo python=3.11.4
   pip install -r requirements.txt

Running the script:

python main.py --url <server_url> --voice <voice>
# Or if you are using authentication
python main.py --url <server_url> --voice <voice> --auth <username:password>

If you do not want to input your microphone and speaker IDs each time, then the use following command with the two IDs you have been choosing:

python main.py --url <server_url> --voice <voice> \
               --input_device <microphone id> \
               --output_device <speaker id>

usage: main.py [-h] --url URL [--auth AUTH] [--debug] [--num_chunks NUM_CHUNKS] [--wave_file_path WAVE_FILE_PATH]
[--voice VOICE] [--vad VAD] [--gpu GPU] [--extra_convert_size EXTRA_CONVERT_SIZE] [--pitch PITCH]
[--crossfade_offset_rate CROSSFADE_OFFSET_RATE] [--crossfade_end_rate CROSSFADE_END_RATE]
[--crossfade_overlap_size CROSSFADE_OVERLAP_SIZE] [--input_device INPUT_DEVICE] [--output_device OUTPUT_DEVICE]

Resemble.AI LiveVC socket sample script. Press Ctrl+C to stop.

options:
  -h, --help            show this help message and exit
  --url URL             URL of the server (required)
  --auth AUTH           ngrok `username:password` for authentication.
  --debug               Enable debug mode for logging.

client parameters:
  --num_chunks NUM_CHUNKS           Number of 2880-frame chunks to send to the server (default: 8).
  --wave_file_path WAVE_FILE_PATH   Path to save the WAV file (default: output.wav).

voice parameters:
  --voice VOICE                                         Name of the voice to use for synthesis.
  --vad VAD                                             VAD level (0: off, 1: low, 2: medium, 3: high) (default: 1).
  --gpu GPU                                             CUDA device ID (default: 0).
  --extra_convert_size EXTRA_CONVERT_SIZE               Amount of context for the server to use (4096, 8192, 16384,
                                                                             32768, 65536, 131072) (default: 4096).
  --pitch PITCH                                         Pitch factor (default: 0).
  --crossfade_offset_rate CROSSFADE_OFFSET_RATE         Crossfade offset rate (0.0 - 1.0) (default: 0.1)
  --crossfade_end_rate CROSSFADE_END_RATE               Crossfade end rate (0.0 - 1.0) (default: 0.9).
  --crossfade_overlap_size CROSSFADE_OVERLAP_SIZE       Crossfade overlap size (default: 2048).

audio device selection. If not specified the user will be provided with a list of devices to choose from.:
  --input_device INPUT_DEVICE, -i INPUT_DEVICE          Index of the input audio device.
  --output_device OUTPUT_DEVICE, -o OUTPUT_DEVICE       Index of the output audio device.

This implementation uses Socket.IO to connect to the server and achieves real-time voice conversion on a M2 Macbook Air client machine, but a lower level websocket implementation can be made as well.

• request_conversion:
  • Type: Emit
  • Description: Sends audio data to the server
  • Data type: AudioData
  • Triggers: on_response
• request_conversion_debug:
  • Type: Emit
  • Description: Identical to request_conversion, but asks the server to return the unconverted audio. This is good for testing the effects of server latency on local audio stitching on clean audio
  • Data type: AudioData
  • Triggers: on_response
• update_model_settings:
  • Type: Emit
  • Description: Sends updated settings to the server
  • Data type: VoiceSettings
  • Triggers: on_message
• get_settings:
  • Type: Emit
  • Description: Requests the current settings dict from the server.
  • Triggers: on_message
• get_voices:
  • Type: Emit
  • Description: Requests a list of available voices from the server.
  • Triggers: on_message
• get_gpus:
  • Type Emit
  • Description: Requests a list of available GPUs from the server.
  • Triggers: on_message
• on_connect:
  • Type: Response
  • Description: Callback for when a connection is established to the server
• on_disconnect:
  • Type: Response
  • Description: Callback for when the connection to the server is disconnected.
• on_response:
  • Type: Response
  • Description: Called when the client receives the audio response from the server.
  • Data type: MessageResponse
• on_message:
  • Type: Response
  • Description: Called when the client receives a message from the server.
  • Data type: MessageResponse

All data sent to and from the server will be in the following data types.

class MessageResponse(TypedDict):
    status: HTTPStatus
    message: str | dict
    endpoint: str
class AudioData(TypedDict):
    timestamp: int      # milliseconds
    audio_data: bytes   # packed little-endian short integers
class VoiceSettings(TypedDict):
    voice: str                 # The name of the voice being used
    crossFadeOffsetRate: float # 0.0 - 1.0
    crossFadeEndRate: float    # 0.0 - 1.0 
    crossFadeOverlapSize: int  # 2048
    extraConvertSize: Literal[4096, 8192, 16384, 32768, 65536, 131072]
    gpu: int    # CUDA device ID
    pitch: float
    vad: Literal[0, 1, 2, 3] # 0: off, 1: low, 2: medium, 3: high

These constants are configured to be exactly what the server expects. They cannot be changed.

• SAMPLERATE = 48000: Sample rate for audio processing.
• CHUNK_SIZE_MULT = 2880: Chunk size multiplier.
• AUDIO_FORMAT = 'int16': Audio format.
• ENDPOINT = '/synthesize': Socket endpoint.

For more information, please refer to the code comments and docstrings within the scripts.

This project is licensed under MIT. See the LICENSE file for details.