Text-to-Speech Latency Benchmark
In the era of large language models (LLMs), Text-to-Speech (TTS) engines are in high demand for use in voice assistants. One of the most critical factors in making conversations with AI agents feel natural and seamless is the response time. A responsive TTS system is essential for delivering a smooth conversational experience, ensuring that the AI's responses are conveyed promptly and without noticeable delays.
This benchmark evaluates the response times of different TTS engines when used in LLM-based voice assistants.
Picovoice Orca Streaming Text-to-Speech is compared against well-known TTS providers:
Methodology
This benchmark simulates user - voice-assistant interactions, by generating LLM responses to user questions
and synthesizing the response to speech as soon as possible.
We sample user queries from a public dataset and feed them to ChatGPT (gpt-3.5-turbo)
using OpenAI Chat Completion API.
ChatGPT generates responses token-by-token, which are passed to different text-to-speech (TTS) engines
to compare their response times.
Engines
All TTS engines listed above support streaming audio output. In addition, Elevenlabs also supports streaming text input using a websocket API. This is done by chunking the text at punctuation marks and sending pre-analyzed text chunks to the engine. Orca Streaming Text-to-Speech supports input text streaming without relying on special language markers. Orca can handle the raw LLM tokens as soon as they are produced.
Dataset
The public taskmaster2 dataset contains text data of goal oriented conversations between a user and an assistant. We randomly select user questions from these example conversations and use them as input to the LLM. The topics of the user queries are diverse and include flight booking, food ordering, hotel booking, movies and music recommendations, restaurant search, and sports. The LLM is prompted to answer the questions like a helpful voice assistant to simulate a real-world user - AI agent interactions. The responses of the LLM have various lengths, from a few words to a few sentences, to cover a wide range of realistic responses.
Metrics
Response times are typically measured with the time-to-first-byte metric, which is the time taken from the moment a
request was sent until the first byte is received.
In the context of assistants we care about the time it takes for the assistant to respond to the user. For LLM-based voice assistants we define:
- Voice Assistant Response Time (
VART): Time taken from the moment the user's request is sent to the LLM, until the TTS engine produces the first byte of speech.
The VART metric is the sum of the following components:
- Time to First Token (
TTFT): Time taken from the moment the user's request is sent to the LLM, until the LLM produces the first byte of text. - First Token to Speech (
FTTS): Time taken from the moment the LLM produces the first text token, until the TTS engine produces the first byte of speech.
The TTFT metric depends on the LLM and possibly network latency if an API is used.
The FTTS metric depends on the capabilities of the TTS engine, and whether it can handle streaming input text,
as well as the token-generation speed of the LLM.
In order to measure the FTTS metric, it is important to keep the LLM behavior constant across all experiments.
We believe the FTTS metric is the most appropriate way to measure the response time of a TTS engine in the context of
voice assistants. This is because it gets closest to the behavior of humans, who can start reading a response as
soon as the first token appears.
Results
The figures below show the response times of each engine by calculating the average over roughly 200 simulated user - voice assistant interactions.
First Token to Speech
Voice Assistant Response Time
Usage
The data and code used to create this benchmark are available on GitHub under the permissive Apache 2.0 license. Detailed instructions for benchmarking individual engines are provided in the following documents: