TTS Decision Guide

Choose the right TTS model for your use case. Includes decision flowcharts, ranked recommendations by scenario, and step-by-step quickstart guides for the top models.

Decision Flowchart

DGX Spark Picks (confirmed or high-confidence)

Preset Voice Models (no cloning needed)

Edge Picks (CPU / Raspberry Pi / Mobile)

Recommendations by Use Case

Voice Assistant / Conversational AI

Goal: Sub-1-second voice-to-voice response with natural conversation flow.

Audiobook / Long-Form Narration

Goal: Highest quality, consistent voice over 10+ minutes, emotion range.

Voice Cloning for Persona Preservation

Goal: Capture a specific person's voice from a recording for ongoing use.

Game / Interactive Media

Goal: Multiple distinct characters, emotional range, real-time.

Multilingual Applications

Goal: High quality across many languages.

The 5-Minute Comparison

Top 10 models ranked by "deployability on DGX Spark"

This ranking prioritizes: (1) confirmed/likely Spark compatibility, (2) voice cloning support, (3) low latency, (4) permissive license.

Quickstart Guides

Quickstart 1: Qwen3-TTS on DGX Spark (Recommended Starting Point)

Quickstart 2: Orpheus 3B via llama.cpp (Most Portable)

Quickstart 3: TADA-1B (Fastest Raw Inference)

Quickstart 4: Chatterbox Turbo (Easiest Setup)

Quickstart 5: Kokoro-82M (Lightest Weight — No Cloning)

Voice Cloning Best Practices

Recording reference audio

The quality of your voice clone depends heavily on the reference audio. Follow these guidelines:

Duration:

• 3 seconds: Minimum viable. Captures basic timbre but misses speaking style
• 10-15 seconds: Sweet spot for most models. Captures timbre + rhythm + style
• 30 seconds: Diminishing returns for most models; some (Fish S2 Pro) benefit
• Over 60 seconds: Can actually degrade quality on some models (Qwen3-TTS)

Recording quality:

• Use a decent microphone (even a modern phone in a quiet room works)
• Record in WAV format, 16-bit, 24 kHz or higher, mono
• Minimize background noise (no music, TV, fans)
• Avoid room echo (record in a carpeted room, not a bathroom)
• Speak naturally — like a conversation, not a voiceover

Content:

• Natural, continuous speech with varied intonation
• Include some questions, statements, and exclamations for prosody range
• Avoid long pauses (> 2 seconds)
• Speech should occupy at least 60% of the total duration
• Provide the accurate transcript if the model requires it

Post-processing:

• Trim silence from beginning and end (leave ~0.5s)
• Normalize to -18 dB LUFS average
• Remove any clicks, pops, or noise artifacts
• Do NOT apply compression, EQ, or effects — models work best with natural audio

Per-model cloning tips

Troubleshooting Common Issues

DGX Spark Specific

Problem: "Torch not compiled with CUDA enabled" or "no kernel image is available"

# Solution: Install PyTorch from the cu130 index
pip install torch==2.9.1 torchaudio==2.9.1 \
  --force-reinstall --index-url https://download.pytorch.org/whl/cu130

Problem: White noise after ~2 seconds ("Blackwell Noise") This affects xTTS-v2, F5-TTS, Kokoro (GPU), and others. It's a numerical instability in FP16/BF16 on sm_121 Tensor Cores.

# Workaround 1: Force CPU inference
device = "cpu"  # in your Python code
 
# Workaround 2: Use a different model that's confirmed working on Spark
# (Qwen3-TTS, VibeVoice-RT, Magpie)

Problem: torchaudio fails to install on aarch64

# Solution 1: Use the cu130 index
pip install torchaudio==2.9.1 --index-url https://download.pytorch.org/whl/cu130
 
# Solution 2: Replace torchaudio with soundfile
pip install soundfile
# Then use: import soundfile as sf; sf.write("out.wav", audio, sr)

pip install flash-attn==2.* --no-build-isolation

Problem: vLLM crashes with CUDA errors

# Solution: Force eager mode
vllm serve model_name --enforce-eager --trust-remote-code

Problem: SGLang sgl-kernel import errors

# Solution: Build from source for sm_121a
cd sglang/sgl-kernel
TORCH_CUDA_ARCH_LIST="12.1a" MAX_JOBS=4 CMAKE_BUILD_PARALLEL_LEVEL=1 \
  python -m build --wheel --no-isolation
pip install --no-deps dist/sgl_kernel*.whl

General TTS Issues

Problem: Generated speech sounds robotic or choppy

• Increase reference audio duration (if using voice cloning)
• Reduce speaking rate (lower rate parameter)
• Try a different model — quality varies significantly between models
• Ensure you're running on GPU (CPU inference can introduce quality artifacts on some models)

Problem: Words are skipped or hallucinated

• Try TADA (zero hallucinations by design)
• Reduce temperature parameter
• Split long text into shorter segments
• Some models hallucinate more at the end of long sequences — limit output length

Problem: Voice clone doesn't sound like the reference

• Ensure reference audio is clean (no background noise, echo)
• Check that reference transcript matches the audio exactly
• Try a longer reference clip (10-15 seconds)
• Some models work better with same-language references
• Reduce cfg_weight or classifier-free guidance scale

Problem: High latency / slow generation

• Check RTF — if > 1.0, generation is slower than real-time
• Enable streaming output to reduce perceived latency
• Use smaller model variant (0.6B instead of 1.7B)
• Apply quantization (INT4 or GGUF Q4)
• Ensure GPU is being used (check nvidia-smi)
• Pre-warm model on startup (run a dummy generation)

This is a companion document to "The Complete Offline TTS Guide (2026 Edition)." Return to the main guide for comprehensive model comparisons and deployment configurations.