Best TTS for Ollama (2026): Adding Voice Output to a Local LLM Setup

Does Ollama Have Built-In Text-to-Speech?

How to Pipe Ollama Output to a Local TTS Engine

• Ollama's API does not know or care what happens to its text output. There is no callback, webhook, or plugin system connecting Ollama to any TTS engine — your code is the only thing joining them.
• Streaming mode ("stream": true) lowers perceived latency by returning tokens as they generate, letting you start synthesizing audio for the first sentence before the model finishes the full reply — useful for interactive voice assistants, more complex to implement than the non-streaming example above.

1. 1
   Start Ollama and pull a model
   Why it matters: Ollama must already be running (`ollama serve`, or the desktop app) with at least one model pulled (`ollama pull llama3.1`) before it can answer requests over its REST API.
2. 2
   Send a prompt to Ollama's REST API
   Why it matters: A POST request to `http://localhost:11434/api/generate` with `"stream": false` returns a single JSON object containing the full reply in its `response` field — simplest to parse for a TTS pipeline, though streaming mode is available for lower time-to-first-audio.
3. 3
   Extract the text and pass it to your TTS engine
   Why it matters: The `response` string is plain text — pass it directly to a TTS engine's CLI over stdin (Piper) or its Python API (Kokoro, XTTS v2, Chatterbox, Bark, or the Coqui TTS toolkit).
4. 4
   Play or save the resulting audio
   Why it matters: Most TTS CLIs and APIs write a `.wav` file directly; for live playback, pipe raw audio to a player like `aplay` (Linux) or use a Python audio library.

# 1. Ask Ollama for a text response (non-streaming, for simplicity)
RESPONSE=$(curl -s http://localhost:11434/api/generate -d '{
  "model": "llama3.1",
  "prompt": "Explain quantum entanglement in two sentences.",
  "stream": false
}' | python3 -c "import sys, json; print(json.load(sys.stdin)['response'])")

# 2. Pipe that text into Piper's CLI to synthesize audio (lowest-resource option)
echo "$RESPONSE" | piper --model en_US-lessac-medium --output_file response.wav

# --- Equivalent Python version, swapping in Kokoro instead of Piper ---
import json
import requests
import soundfile as sf
from kokoro_onnx import Kokoro

reply = requests.post(
    "http://localhost:11434/api/generate",
    json={"model": "llama3.1", "prompt": "Explain quantum entanglement in two sentences.", "stream": False},
).json()["response"]

kokoro = Kokoro("kokoro-v1.0.onnx", "voices-v1.0.bin")
samples, sample_rate = kokoro.create(reply, voice="af_heart")
sf.write("response.wav", samples, sample_rate)

Which TTS Engine Pairs Best with Ollama?

Who Should Use Which Engine?

• 🏆 Best overall for an Ollama pairing: Piper — lowest resource cost, real-time on CPU, simplest to pipe into a shell script or a Python subprocess call.
• Best for higher audio quality at a similar footprint: Kokoro — still small enough to run without a GPU, with noticeably better perceived speech quality than Piper per its own release benchmarks.
• Best for voice cloning, commercial use allowed: Chatterbox — MIT-licensed, clones a voice from about 5 seconds of reference audio, needs a GPU alongside Ollama for real-time use.
• Best for voice cloning, non-commercial or research use: XTTS v2 — clones a voice from 6 seconds of audio across 17 languages, but its CPML license blocks commercial use without a separate agreement — see PromptQuorum's XTTS v2 license breakdown.
• Best for expressive non-speech audio, not as a primary voice: Bark — laughter, sighs, and simple ambient sound from text prompts alone, but its repository has had no commits since April 5, 2024, so do not depend on it for a maintained production pipeline.
• 🧭 Raspberry Pi or other CPU-only hardware, running Ollama with a small model → Piper. Nothing else in this guide is confirmed to run in real time without a GPU.
• 🧭 Desktop or server with a spare GPU alongside Ollama, want a cloned voice, and need commercial rights → Chatterbox.
• 🧭 Desktop or server with a spare GPU, research or personal project, want the highest cloning quality → XTTS v2.
• 🧭 Want a single toolkit that can load several different models over time (including XTTS v2) → Coqui TTS toolkit instead of installing each model's dependencies separately.

When Not to Use Any of These

• ❌ If you need dozens of highly polished, emotionally expressive voices out of the box — a managed cloud API such as ElevenLabs offers a larger curated voice library and more expressive controls than any of the models here; see PromptQuorum's ElevenLabs vs. local TTS comparison for the trade-offs.
• ❌ If your hardware cannot spare RAM or VRAM beyond what Ollama already uses — running Ollama and a GPU-hungry TTS engine like XTTS v2 or Bark on the same modest GPU can starve both; drop to Piper or Kokoro, or move TTS to a second machine.
• ❌ If you need a shipped commercial product and have not independently confirmed a license — XTTS v2's CPML is explicitly non-commercial, and Coqui AI, the company behind it, shut down its paid licensing services in December 2023; verify licensing terms yourself before shipping any of these engines in a paid product.
• ❌ If you are cloning a real person's voice without their consent — this raises consent and impersonation concerns independent of any engine's license, in personal and commercial use alike.

Frequently Asked Questions

Verdict

Sources

• Ollama GitHub Issue #11021 — the feature request for native TTS support, closed as a duplicate of an earlier, still-open issue.
• Ollama REST API documentation — the /api/generate endpoint used in this guide's pipeline examples.
• maudoin/ollama-voice on GitHub — a community project chaining Whisper, Ollama, and pyttsx3; 378 stars at the time of writing.
• Kokoro-82M model card on Hugging Face — parameter count, license, and architecture.
• XTTS v2 on Hugging Face — cloning requirements and license reference.
• Chatterbox on GitHub — Resemble AI's open-source voice-cloning model, license, and usage examples.
• Bark on GitHub — Suno's generative audio model, license, and commit history.

Related Reading

• Piper TTS Review — the fast, CPU-only local TTS engine this guide ranks first for Ollama pairings.
• XTTS v2 Review — the full write-up on voice cloning, its non-commercial license, and real usage commands.
• Coqui TTS Review — the toolkit that runs XTTS v2 and other models through one Python API.
• Bark TTS Review — the expressive, non-speech-audio model, including its maintenance status.
• Local TTS & Voice Cloning Licenses — a full licensing comparison across local TTS engines.
• ElevenLabs vs. Piper vs. XTTS v2 — the cloud-vs-local comparison for readers deciding whether to self-host at all.
• Local Voice Assistant Build Guide — a full Whisper-plus-Ollama-plus-Piper assistant build, for readers who also need speech recognition on the input side.