​
Overview

Create custom voice models to generate consistent, high-quality speech. You can create models through our web interface or programmatically via API.

​
Web Interface

The easiest way to create a voice model:
1

Go to Fish Audio

Visit fish.audio and log in
2

Navigate to Models

Click on “Models” in your dashboard
3

Click Create Model

Select “Create New Model”
4

Upload Your Audio

Add 2 or more voice samples (30-45 seconds each)
5

Configure Settings

Choose privacy settings and training options
6

Start Training

Click “Create” and wait for processing

​
Using the API

​
Python SDK

First, install the SDK: 
pip install fish-audio-sdk
Then create a model: 
from fish_audio_sdk import Session

# Initialize session with your API key
session = Session("your_api_key")

# Create the model
model = session.create_model(
    title="My Voice Model",
    description="Custom voice for storytelling",
    voices=[
        voice_file1.read(),
        voice_file2.read()
    ],
    cover_image=image_file.read()  # Optional
)

print(f"Model created: {model.id}")

​
Direct API

Create models directly using the REST API: 
import requests

response = requests.post(
    "https://api.fish.audio/model",
    files=[
        ("voices", open("sample1.mp3", "rb")),
        ("voices", open("sample2.wav", "rb"))
    ],
    data=[
        ("title", "My Voice Model"),
        ("description", "Custom voice model"),
        ("visibility", "private"),
        ("type", "tts"),
        ("train_mode", "fast"),
        ("enhance_audio_quality", "true")
    ],
    headers={
        "Authorization": "Bearer YOUR_API_KEY"
    }
)

result = response.json()
print(f"Model ID: {result['id']}")

​
Model Settings

​
Required Parameters

ParameterDescriptionOptions
titleName of your modelAny text
voicesAudio samples (minimum 2)MP3, WAV, M4A files

​
Optional Parameters

ParameterDescriptionOptions
descriptionModel descriptionAny text
visibilityWho can use your modelprivate, public
train_modeTraining speed/qualityfast, balanced, quality
enhance_audio_qualityRemove background noisetrue, false
textsTranscripts of audio samplesMust match number of audio files
cover_imageModel thumbnailJPG, PNG image

​
Audio Requirements

​
Quality Guidelines

Minimum Requirements:
  • At least 2 audio samples
  • 30-45 seconds per sample
  • Single speaker only
  • Consistent voice throughout
Best Practices:
  • Use 3-5 diverse samples
  • Include different emotions and tones
  • Record in a quiet environment
  • Maintain steady volume

​
Training Modes

​
Fast Mode

  • Training Time: 5-10 minutes
  • Best For: Quick prototypes
  • Quality: Good for basic use

​
Balanced Mode

  • Training Time: 15-30 minutes
  • Best For: Most use cases
  • Quality: Excellent balance

​
Quality Mode

  • Training Time: 45-60 minutes
  • Best For: Professional projects
  • Quality: Maximum accuracy

​
Adding Transcripts

Including text transcripts improves model quality: 
response = requests.post(
    "https://api.fish.audio/model",
    files=[
        ("voices", open("hello.mp3", "rb")),
        ("voices", open("world.wav", "rb"))
    ],
    data=[
        ("title", "Enhanced Model"),
        ("texts", "Hello, this is my first recording."),
        ("texts", "Welcome to the world of AI voices."),
        # ... other parameters
    ],
    headers={"Authorization": "Bearer YOUR_API_KEY"}
)
Text transcripts must match the exact number of audio files. If you provide 3 audio files, you must provide exactly 3 text transcripts.

​
Checking Model Status

After creating a model, check its training status: 
# Using SDK
model_status = session.get_model(model_id)
print(f"Status: {model_status.status}")

# Using API
response = requests.get(
    f"https://api.fish.audio/model/{model_id}",
    headers={"Authorization": "Bearer YOUR_API_KEY"}
)
status = response.json()["status"]
Model statuses:
  • pending: In queue for training
  • training: Currently being processed
  • completed: Ready to use
  • failed: Training unsuccessful

​
Using Your Model

Once training is complete: 
# Generate speech with your model
response = requests.post(
    "https://api.fish.audio/v1/tts",
    json={
        "text": "Hello from my custom voice!",
        "model_id": model_id,
        "format": "mp3"
    },
    headers={"Authorization": "Bearer YOUR_API_KEY"}
)

# Save the audio
with open("output.mp3", "wb") as f:
    f.write(response.content)

​
Troubleshooting

​
Common Issues

Model training fails:
  • Check audio quality and format
  • Ensure single speaker in all samples
  • Verify files are not corrupted
Poor voice quality:
  • Add more diverse audio samples
  • Enable audio enhancement
  • Use higher quality training mode
Long training times:
  • Normal for quality mode
  • Check API status page for delays
  • Consider using balanced mode

​
Best Practices

  1. Start Simple: Begin with 2-3 samples in fast mode to test
  2. Iterate: Refine with more samples and quality mode
  3. Document: Keep track of which samples work best
  4. Test Thoroughly: Try different texts and emotions
  5. Privacy First: Keep personal models private

​
Support

Need help creating models?