loudness-tts

Fast, natural text to speech on your own computer or server. Free, offline after the first voice download, no API key. By Loudness.ai.

• 1,683 ranked voices in 32 language variants: every Piper voice that allows commercial use, scored for naturalness and clarity in our open study. loudness-tts picks the best voice for your language automatically and shows each voice's licence and credit line.
• Reads like a person: prices, dates, times, units, phone numbers, abbreviations and Roman numerals become words (English and Spanish), with natural pauses, calmer delivery and click-free joins (how and how well).
• Ready to publish: loudness normalized to −16 LUFS with true peak at or below −1.5 dBTP, word timings, SRT and VTT subtitles.
• Fast on a CPU: 16 to 21 times faster than real time with medium voices on an Apple M4, first streamed sentence in about 0.1 s. No PyTorch, no GPU, no FFmpeg needed (FFmpeg only adds MP3/Opus/AAC/FLAC output).
• Works where you work: Python library, command line, an OpenAI-compatible speech server, Docker, and a JavaScript/TypeScript edition for Node, Bun, Deno and browsers that produces the same speech.

pip install loudness-tts
loudness-tts "Dr. Smith paid $5.50 on 4/5/2026." -o speech.wav --srt speech.srt

The first run downloads the best English (US) voice (about 78 MB, checked against its published checksum). Try voices in the browser first: Piper Voice Explorer · Loudness.ai Text to Speech.

Python

import loudness_tts

loudness_tts.speak("Hello! The meeting starts at 9:30 AM.", "hello.wav")          # best English (US) voice

speech = loudness_tts.synthesize("Llegamos el 15/10/2026 a las 10:30.", language="es-ES")
speech.save("es.wav")            # 16-bit WAV (speech.save(path, bit_depth=24) for 24-bit)
speech.save_srt("es.srt")        # subtitles; speech.vtt() for web players
print(speech.spoken)             # what the voice actually read
print(speech.voice.credit)       # credit line, when the voice's licence asks for one

for chunk in loudness_tts.stream("A long text is spoken sentence by sentence, so playback starts at once."):
    player.write(chunk.pcm())    # 16-bit mono PCM at chunk.sample_rate

loudness_tts.voices(language="en-US", sex="female")[:3]                         # best English (US) female voices

For services, keep one engine so voices stay loaded:

from loudness_tts import TTS

tts = TTS(language="en-US", threads=2, cache_size=4)
tts.preload()                                         # download and load now, not on the first request
speech = tts.synthesize(text, voice="en_US-libritts-high-s118", speed=1.1, pause="short", timeout=30)

Word timings are estimated from the text by default; pip install "loudness-tts[timings]" makes them exact (taken from the voice model's own sound durations).

Command line

loudness-tts "Text to speak" -o out.wav [--voice ID] [--language es-ES] [--speed 1.1] [--srt out.srt] [--vtt out.vtt]
echo "Text from a pipe" | loudness-tts - > out.wav
loudness-tts voices --language de --min-score 75
loudness-tts languages
loudness-tts download --language de-DE          # for offline machines and Docker images
loudness-tts doctor --synthesize                # check the installation

OpenAI-compatible server

loudness-tts serve --port 8880                  # add --host 0.0.0.0 --api-key SECRET to serve other machines

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8880/v1", api_key="not-needed")
client.audio.speech.create(model="tts-1", voice="nova", input="Hello from my own server!",
                          response_format="wav").write_to_file("hello.wav")

Apps that let you set an OpenAI-compatible text-to-speech URL, such as Open WebUI, LibreChat and AnythingLLM, can use it directly: base URL http://127.0.0.1:8880/v1, any model name, a voice id or an OpenAI voice name. response_format pcm streams 24 kHz audio while it is made; mp3, opus, aac and flac need FFmpeg. POST /v1/tts returns JSON with the audio, word timings, SRT and VTT. Text is never logged or stored.

Docker

docker run -p 8880:8880 -v loudness-tts:/data ghcr.io/loudness-ai/loudness-tts

Voices and licences

Every voice in the catalogue allows commercial use. About 1,600 ask for a credit line (CC BY, CC BY-SA, Apache 2.0), 51 need nothing, two are copyleft and 17 have unclear terms; loudness-tts never picks those last two groups on its own. voice.licence, voice.use and voice.credit tell you what applies. Voice files come from their original sources (the official Piper voices at a pinned revision, or the community author) and are verified before use. The ranking and its method: Open Piper TTS Voices Ranked, with the data on Hugging Face.

Privacy

Speech is made on your machine. The only network requests are voice downloads (one per voice, from Hugging Face or the voice author). Set LOUDNESS_TTS_OFFLINE=1 to forbid them; LOUDNESS_TTS_HOME moves the voice cache.

Licence and credits

GPL-3.0-or-later, because it builds on Piper and espeak-ng. The text preparation layer (loudness_tts/textprep.py) is Apache-2.0 and can be reused on its own. Voices keep their own licences. If loudness-tts helps your project, a link to loudness.ai is appreciated (not required).

@software{loudness_tts,
  author = {Shehzad, Asad},
  title = {loudness-tts: natural text to speech on your own computer or server},
  publisher = {Loudness.ai},
  year = {2026},
  url = {https://github.com/Loudness-ai/loudness-tts}
}

More from Loudness.ai: free online audio tools · Audio Kit (local noise removal and loudness normalization) · Research & Benchmarks · Hugging Face