​
1. Use cases

ext-to-Speech (TTS) models are AI models that convert text information into spoken output. These models generate natural and expressive speech from input text, suitable for various use cases:
  • Providing audio readings for blog articles
  • Generating multilingual speech content
  • Supporting real-time streaming audio output

​
2. API usage guide

  • Endpoint: /audio/speech. For detailed usage, refer to the API documentation
  • Main request parameters:：
    • model: The model used for speech synthesis, with a list of supported models。
    • input: The text content to be converted into audio.
    • voice: The reference voice, supporting system-predefined voices、user-predefined voices、user-dynamic voices. Detailed parameters: Refer toCreating a text-to-speech request。
    • speed: Controls the audio speed, a float type with a default value of 1.0 and a selectable range of [0.25, 4.0].
    • gain: Audio gain in dB, controls the volume of the audio, a float type with a default value of 0.0 and a selectable range of [-10, 10].
    • response_format: Controls the output format, supporting mp3, opus, wav, and pcm formats. Different output formats result in different sampling rates.
    • sample_rate: Controls the output sampling rate, with different default values and selectable ranges depending on the output format:
      • opus: Currently supports only 48000 Hz.
      • wav, pcm: Supports (8000, 16000, 24000, 32000, 44100), default is 44100.
      • mp3: Supports (32000, 44100), default is 44100.

​
2.1 System-predefined voices:

Currently, the system provides the following 8 voice options:
  • male voices:
  • female voice:
    • steady female voice: anna
    • passionate female voice: bella
    • gentle female voice: claire
    • cheerful female voice: diana
online listeningof the above audio. When using the system-predefined voices in the request, you need to prepend the model name, such as: FunAudioLLM/CosyVoice2-0.5B:alex indicates the alex voice from the FunAudioLLM/CosyVoice2-0.5B model.

​
2.2 User-predefined voices:

Note: Using user-predefined voices requires real-name authentication.
To ensure the quality of the generated voice, it is recommended that users upload a voice sample that is 8 to 10 seconds long, with clear pronunciation and no background noise or interference.

​
2.2.1 Upload user-predefined voices using base64 encoding format

import requests
import json

url = "https://api.siliconflow.cn/v1/uploads/audio/voice"
headers = {
    "Authorization": "Bearer your-api-key", # 从 https://cloud.siliconflow.cn/account/ak 获取
    "Content-Type": "application/json"
}
data = {
    "model": "FunAudioLLM/CosyVoice2-0.5B", # model name
    "customName": "your-voice-name", # user-predefined voices name
    "audio": "data:audio/mpeg;base64,SUQzBAAAAAAAIlRTU0UAAAAOAAADTGF2ZjYxLjcuMTAwAAAAAAAAAAAAAAD/40DAAAAAAAAAAAAASW5mbwAAAA8AAAAWAAAJywAfHx8fKioqKio1NTU1Pz8/Pz9KSkpKVVVVVVVfX19fampqamp1dXV1f39/f3+KioqKlZWVlZWfn5+fn6qqqqq1tbW1tb+/v7/KysrKytXV1dXf39/f3+rq6ur19fX19f////", # 参考音频的 base64 编码
    "text": "在一无所知中, 梦里的一天结束了，一个新的轮回便会开始" # 参考音频的文字内容
}

response = requests.post(url, headers=headers, data=json.dumps(data))

# 打印响应状态码和响应内容
print(response.status_code)
print(response.json())  # 如果响应是 JSON 格式
The URI field in the response is the ID of the custom voice. Users can use this ID as the voice parameter in subsequent requests. 
{'uri': 'speech:your-voice-name:cm04pf7az00061413w7kz5qxs:mjtkgbyuunvtybnsvbxd'}
In the request, use the user-predefined voices as indicated.

​
2.2.2 Upload user-predefined voices through a file

import requests

url = "https://api.siliconflow.cn/v1/uploads/audio/voice"
headers = {
    "Authorization": "Bearer your-api-key" # 从 https://cloud.siliconflow.cn/account/ak 获取
}
files = {
    "file": open("/Users/senseb/Downloads/fish_audio-Alex.mp3", "rb") # 参考音频文件
}
data = {
    "model": "FunAudioLLM/CosyVoice2-0.5B", # 模型名称
    "customName": "your-voice-name", # 参考音频名称
    "text": "在一无所知中, 梦里的一天结束了，一个新的轮回便会开始" # 参考音频的文字内容
}

response = requests.post(url, headers=headers, files=files, data=data)

print(response.status_code)
print(response.json())
The URI field in the response is the ID of the custom voice. Users can use this ID as the voice parameter in subsequent requests. 
{'uri': 'speech:your-voice-name:cm04pf7az00061413w7kz5qxs:mjtkgbyuunvtybnsvbxd'}
In the request, use the user-predefined voices as indicated.

​
2.3 Get the list of user-dynamic voices

import requests
url = "https://api.siliconflow.cn/v1/audio/voice/list"

headers = {
    "Authorization": "Bearer your-api-key" # 从https://cloud.siliconflow.cn/account/ak获取
}
response = requests.get(url, headers=headers)

print(response.status_code)
print(response.json())
The URI field in the response is the ID of the custom voice. Users can use this ID as the voice parameter in subsequent requests. 
{'uri': 'speech:your-voice-name:cm04pf7az00061413w7kz5qxs:mjtkgbyuunvtybnsvbxd'}
In the request, use the user-predefined voices as indicated.

​
2.4 Use user-dynamic voices

Note: Using user-predefined voices requires real-name authentication.
Use the user-dynamic voices in the request as indicated.

​
2.5 Delete user-dynamic voices

import requests

url = "https://api.siliconflow.cn/v1/audio/voice/deletions"
headers = {
    "Authorization": "Bearer your-api-key",
    "Content-Type": "application/json"
}
payload = {
    "uri": "speech:your-voice-name:cm02pf7az00061413w7kz5qxs:mttkgbyuunvtybnsvbxd"
}

response = requests.request("POST", url, json=payload, headers=headers)

print(response.status_code)
print(response.text)
The URI field in the response is the ID of the custom voice. Users can use this ID as the voice parameter in subsequent requests.

​
3. List of supported models

Note: The supported TTS models may be subject to change. Please filter by the “Speech” tag on the 「Models」to obtain the current list of supported models.
Charging method: Charged based on the number of UTF-8 bytes corresponding to the input text. Online byte counter demo.

​
3.1 FunAudioLLM/CosyVoice2-0.5B series models

  • Cross-Language text-to-speech: Achieve text-to-speech synthesis across different languages, including Chinese, English, Japanese, Korean, Chinese dialects (Cantonese, Sichuanese, Shanghai dialect, Zhengzhou dialect, Changsha dialect, Tianjin dialect)
  • Emotional control: Support the generation of speech with various emotional expressions, including happiness, excitement, sadness, anger, etc.
  • Fine-grained control: Control the emotional and rhythmic aspects of the generated speech through rich text or natural language.

​
4. Best practices for reference audio

Providing high-quality reference audio samples can enhance the cloning effect.

​
4.1 Audio quality guidelines

  • Single speaker only
  • Clear articulation, steady volume, tone, and emotion
  • Brief pauses (0.5 seconds recommended)
  • Ideal: No background noise, professional recording quality, no room echo
  • Suggested duration: 8 to 10 seconds

​
4.2 File formats

  • Supported formats: mp3, wav, pcm, opus
  • Recommend using mp3 with a bitrate above 192kbps to avoid quality loss
  • Additional benefits of uncompressed formats (e.g., WAV) are limited

​
5. Usage examples

​
5.1 Use system-predefined voices

from pathlib import Path
from openai import OpenAI

speech_file_path = Path(__file__).parent / "siliconcloud-generated-speech.mp3"

client = OpenAI(
    api_key="您的 APIKEY", # 从 https://cloud.siliconflow.cn/account/ak 获取
    base_url="https://api.siliconflow.cn/v1"
)

with client.audio.speech.with_streaming_response.create(
  model="FunAudioLLM/CosyVoice2-0.5B", # 支持 fishaudio / GPT-SoVITS / CosyVoice2-0.5B 系列模型
  voice="FunAudioLLM/CosyVoice2-0.5B:alex", # 系统预置音色
  # 用户输入信息
  input="你能用高兴的情感说吗？<|endofprompt|>今天真是太开心了，马上要放假了！I'm so happy, Spring Festival is coming!",
  response_format="mp3" # 支持 mp3, wav, pcm, opus 格式
) as response:
    response.stream_to_file(speech_file_path)

​
5.2 Use user-predefined voices

from pathlib import Path
from openai import OpenAI

speech_file_path = Path(__file__).parent / "siliconcloud-generated-speech.mp3"

client = OpenAI(
    api_key="您的 APIKEY", # 从 https://cloud.siliconflow.cn/account/ak 获取
    base_url="https://api.siliconflow.cn/v1"
)

with client.audio.speech.with_streaming_response.create(
  model="FunAudioLLM/CosyVoice2-0.5B", # 支持 fishaudio / GPT-SoVITS / CosyVoice2-0.5B 系列模型
  voice="speech:your-voice-name:cm02pf7az00061413w7kz5qxs:mttkgbyuunvtybnsvbxd", # 用户上传音色名称，参考
  # 用户输入信息
  input=" 请问你能模仿粤语的口音吗？< |endofprompt| >多保重，早休息。",
  response_format="mp3"
) as response:
    response.stream_to_file(speech_file_path)

​
5.3 Use user-dynamic voices

from pathlib import Path
from openai import OpenAI
client = OpenAI()

speech_file_path = Path(__file__).parent / "siliconcloud-generated-speech.mp3"

client = OpenAI(
    api_key="您的 APIKEY", # 从 https://cloud.siliconflow.cn/account/ak 获取
    base_url="https://api.siliconflow.cn/v1"
)

with client.audio.speech.with_streaming_response.create(
  model="FunAudioLLM/CosyVoice2-0.5B", 
  voice="", # 此处传入空值，表示使用动态音色
  # 用户输入信息
  input="  [laughter]有时候，看着小孩子们的天真行为[laughter]，我们总会会心一笑。",
  response_format="mp3",
  extra_body={"references":[
        {
            "audio": "https://sf-maas-uat-prod.oss-cn-shanghai.aliyuncs.com/voice_template/fish_audio-Alex.mp3", # 参考音频 url。也支持 base64 格式
            "text": "在一无所知中, 梦里的一天结束了，一个新的轮回便会开始", # 参考音频的文字内容
        }
    ]}
) as response:
    response.stream_to_file(speech_file_path)