Fish-Speech 1.4

Training compute

1.9×10²¹ FLOP

Published

Nov 9, 2024

Fish-Speech 1.4 is an AI model developed by Fish Audio (United States), first published in November 2024. It works in the speech domain, on tasks such as speech synthesis and text-to-speech (tts).

Training it took an estimated 1.9×10²¹ FLOP of compute (estimation method: hardware). It was trained on roughly 500B datapoints. Training ran on 8 NVIDIA H100 SXM5 80GB,NVIDIA GeForce RTX 4090 for about 168 hours.

Access: Open weights (non-commercial). Its weights are openly available. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Fish Audio
Country of organization: United States
Domain: Speech
Task: Speech synthesis, Text-to-speech (TTS)
Training compute: 1.9×10²¹ FLOP
Compute estimation method: Hardware
Dataset size: 500B
Training hardware: NVIDIA H100 SXM5 80GB, NVIDIA GeForce RTX 4090
Chips used: 8
Training time: 168 h
Model accessibility: Open weights (non-commercial)
Open weights: Yes
Epoch confidence: Confident