Parler-TTS Self-Hosted

Parler-TTS generates speech controlled by natural-language descriptions of the voice and delivery style – “a calm female voice reading slowly” or “a confident male narrator”. On our dedicated GPU hosting it runs comfortably on any GPU with 8+ GB.

Contents

• Variants
• Deployment
• Prompting voices
• Throughput

Variants

Two practical sizes in 2026:

• Parler-TTS Mini (~600M params, ~1.5 GB VRAM)
• Parler-TTS Large (~2.2B params, ~5 GB VRAM)

Deployment

from parler_tts import ParlerTTSForConditionalGeneration
from transformers import AutoTokenizer
import soundfile as sf

model = ParlerTTSForConditionalGeneration.from_pretrained("parler-tts/parler-tts-large-v1").to("cuda")
tok = AutoTokenizer.from_pretrained("parler-tts/parler-tts-large-v1")

description = "A calm female speaker with a natural English accent and moderate pace."
prompt = "Hello, and welcome to GigaGPU's audio demo."

desc_ids = tok(description, return_tensors="pt").input_ids.to("cuda")
prompt_ids = tok(prompt, return_tensors="pt").input_ids.to("cuda")

audio = model.generate(input_ids=desc_ids, prompt_input_ids=prompt_ids)
sf.write("out.wav", audio[0].cpu().numpy().squeeze(), model.config.sampling_rate)

Prompting Voices

The description controls voice characteristics. Effective description dimensions:

• Gender: “female”, “male”
• Pace: “slowly”, “at moderate pace”, “quickly”
• Pitch: “high-pitched”, “deep”
• Tone: “calm”, “excited”, “serious”
• Accent: “American”, “British”, “Australian”

For brand consistency, standardise a single description across your app and reuse it.

Throughput

Parler-TTS Large on a 4060 Ti:

• Short sentence (~10 words): 1-2 seconds generation
• Paragraph (~50 words): 5-8 seconds generation
• Real-time factor: roughly 0.3x (faster than playback)

Compare Fish Speech, MeloTTS, and Bark.