"""The loudkit demo Space: hear twenty voices, speak your own text, clone your own.
ZeroGPU bills GPU time to the *visitor*, not to the owner: an anonymous visitor
gets about two minutes a day, a signed-in free account about five. A demo whose
first click spends that budget is one most people bounce off before they have
heard anything at all. So the Listen tab is twenty pre-rendered files served
straight out of this repo — no GPU, no queue, no quota — and the GPU is spent
only on what a visitor types or records.
The engine and the enroller are both built on `cuda` at module level, which is
what ZeroGPU asks for: CUDA transfers are optimised for start-up placement, and
lazy-loading inside a `@spaces.GPU` function is explicitly discouraged. Each
decorated call then runs in a freshly forked, short-lived process, which is also
why there is no `torch.compile` and no CUDA graph capture here: both pay their
cost once per process and would never amortise.
Cloning is exposed, which the CPU scaffold this replaces deliberately did not do.
The reasoning that kept it out was about consent, not about capability, so the
consent is built into the shape of the tab rather than written beside it: the
microphone is the default path, an upload is secondary and gated on an explicit
confirmation, and neither recording outlives the request that carried it.
"""
from __future__ import annotations
import contextlib
import dataclasses
import hashlib
import json
import os
import tempfile
from pathlib import Path
# Before torch, and before anything that imports torch. The module installs the
# CUDA emulation that lets a module-level `.to("cuda")` succeed on a machine
# that has no GPU attached yet.
import spaces
import gradio as gr
import numpy as np
import loudkit as lk
from loudkit.backends.torch_backend import build_torch_enroller
from loudkit.hub import resolve_enrollment_checkpoint, resolve_voice_encoder
REPO = "loudreader/loudr-1"
DEVICE = "cuda"
HERE = Path(__file__).parent
# The CPU scaffold capped text at 300 characters because CPU synthesis ran at
# roughly a tenth of real time. On a GPU the cap is about the visitor's daily
# quota instead, which is a far looser bound: 1000 characters is ~70 s of speech.
MAX_CHARS = 1_000
MAX_CLONE_CHARS = 400
MAX_PROBE_CHARS = 200
# The enroller refuses anything over 30 s and wants 5 to 10. The prompt is built
# from the first 10 s; the speaker embedding reads whatever else is there, so a
# little past the prompt window is useful and 20 s stays clear of the refusal.
ENROLL_SECONDS = 20.0
DOCS = "https://github.com/loudreader/loudkit"
IDENTITY_CONTRACT = f"{DOCS}/blob/main/docs/reference/IDENTITY-CONTRACT.md"
RESPONSIBLE_USE = f"{DOCS}/blob/main/RESPONSIBLE_USE.md"
ROSTER = json.loads((HERE / "voices.json").read_text(encoding="utf-8"))
BY_NAME = {entry["name"]: entry for entry in ROSTER}
ORDERED = sorted(ROSTER, key=lambda e: (e["language"], e["name"]))
VOICE_CHOICES = [(f"{e['name']} · {e['language']} ({e['gender']})", e["name"]) for e in ORDERED]
# English leads. A visitor who does not read the other nine should not have to
# hunt for the one they do, and alphabetical order put Danish first.
FIRST_LANGUAGE = "en"
_LANG_OF = {e["language_id"]: e["language"] for e in ROSTER}
LANGUAGE_FILTER = [(_LANG_OF[FIRST_LANGUAGE], FIRST_LANGUAGE)] + [
(name, code) for code, name in sorted(_LANG_OF.items(), key=lambda kv: kv[1])
if code != FIRST_LANGUAGE
]
def voices_in(language: str) -> list[tuple[str, str]]:
return [
(f"{e['name']} ({e['gender']})", e["name"])
for e in ORDERED
if e["language_id"] == language
]
FIRST_VOICES = voices_in(FIRST_LANGUAGE)
FIRST_VOICE = FIRST_VOICES[0][1]
# Reference clips a visitor can clone without recording anything. Every one was
# donated for building TTS voices, so these need no consent box; the roster
# carries the licence and the donor's own words beside each.
CLONE_EXAMPLES = ["joe", "kathleen", "ines", "gosia"]
EXAMPLE_CHOICES = [("Nothing selected", "")] + [
(f"{n} · {BY_NAME[n]['language']}", n) for n in CLONE_EXAMPLES
]
# --------------------------------------------------------------------------
# Module-level model placement, per the ZeroGPU contract.
# --------------------------------------------------------------------------
engine = lk.load(REPO, device=DEVICE)
# Voice profiles are numpy, not torch, so they are device-agnostic and cost a
# few hundred kilobytes each. Loading all twenty up front means switching voice
# in the Speak tab never blocks on a download.
PROFILES = {entry["name"]: lk.voice(entry["name"], repo=REPO) for entry in ROSTER}
# Enrollment reads the other half of the release: the speech tokenizer and the
# speaker encoder, which synthesis never touches, plus the utterance voice
# encoder that sits beside both. `lk.enroll()` builds this per call by design;
# a Space would pay the load on every clone, so it is built once here instead.
enroller = build_torch_enroller(
str(resolve_enrollment_checkpoint(REPO)),
device=DEVICE,
voice_encoder_weights=str(resolve_voice_encoder(REPO)),
)
FINGERPRINT = engine.algorithm.fingerprint()
_LANGUAGE_NAMES = {e["language_id"]: e["language"] for e in ROSTER}
LANGUAGE_CHOICES = [("Follow the voice", "")] + [
(f"{_LANGUAGE_NAMES.get(code, code)} ({code})", code) for code in lk.languages()
]
# --------------------------------------------------------------------------
# Helpers
# --------------------------------------------------------------------------
def _sha256_audio(audio: np.ndarray) -> str:
"""Hash the waveform, not the file.
`Result.save` appends a C2PA manifest carrying a wall-clock creation time,
which the library itself calls the one byte range in which two identical
renders may legitimately differ. Hashing the saved WAV would therefore print
two different digests for two identical renders and read as a determinism
failure. The waveform is what the identity contract makes its promise about,
so the waveform is what gets hashed.
"""
return hashlib.sha256(np.ascontiguousarray(audio, dtype=np.float32).tobytes()).hexdigest()
def _write(result: lk.Result, *, voice: str, language: str) -> str:
out = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
out.close()
# Provenance on: the manifest carries the fingerprint, the recipe and the
# seed, which is the machine-readable marking a synthetic-speech demo should
# be handing out by default.
result.save(out.name, voice=voice, language=language)
return out.name
def _stats(result: lk.Result) -> str:
seconds = len(result.audio) / result.sample_rate
return (
f"**{seconds:.1f} s of audio.** {result.timings.describe(seconds)}\n\n"
f"`algo[{result.algorithm_fingerprint}]` · seed `{result.seed}` · "
f"speed `{result.speed:g}x` · {result.sample_rate} Hz"
)
def _estimate(text: str, *, passes: int = 1, overhead: float = 15.0) -> int:
"""Seconds of GPU to ask for.
Speech runs at roughly 14 characters a second, and the render is asked to
keep up with better than real time; the overhead covers the process fork and
the first real CUDA touch. Asking for too much costs queue priority but not
quota, which is charged on effective duration, so this leans generous.
"""
audio_seconds = len((text or "").strip()) / 14.0
return int(min(180.0, overhead + passes * max(4.0, audio_seconds * 0.9)))
def _check(text: str, limit: int) -> str:
text = (text or "").strip()
if not text:
raise gr.Error("Type something to say.")
if len(text) > limit:
raise gr.Error(f"Keep it under {limit:,} characters here. The library itself takes 10,000.")
return text
# --------------------------------------------------------------------------
# Listen. No GPU: these files were rendered ahead of time and ship in the repo.
# --------------------------------------------------------------------------
def listen(name: str):
entry = BY_NAME[name]
sample, reference, source = entry["sample"], entry["reference"], entry["source"]
lines = [
f"### {entry['name']}. {entry['language']} ({entry['gender']}).",
"",
f"> {sample['text']}",
"",
f"From *{sample['work']}*, seed `{sample['seed']}`.",
"",
f"- Reference recording: {reference['duration_s']:.1f} s, {reference['construction']}.",
f"- Source: [{source['name']}]({source['url']}), {source['license']}.",
f"- Consent: {source['consent']}.",
]
similarity = entry.get("speaker_similarity")
if similarity is not None:
lines.append(f"- Speaker similarity to the reference: {similarity:.3f}.")
lines.append(f"- Voice profile: `{entry['profile']['hf_path']}`.")
return (
str(HERE / sample["audio"]),
str(HERE / reference["public_preview"]),
"\n".join(lines),
)
ROSTER_TABLE = [
[
entry["name"],
entry["language"],
entry["gender"],
entry["source"]["license"],
f"{entry['speaker_similarity']:.3f}" if entry.get("speaker_similarity") is not None else "",
]
for entry in ORDERED
]
# --------------------------------------------------------------------------
# Speak. GPU.
# --------------------------------------------------------------------------
def _speak_duration(text, name, language, seed, speed):
return _estimate(text, overhead=15.0)
@spaces.GPU(duration=_speak_duration)
def speak(text: str, name: str, language: str, seed: float, speed: float):
text = _check(text, MAX_CHARS)
result = engine.synthesize_long(
text,
PROFILES[name],
seed=int(seed),
language=language or None,
speed=float(speed),
)
label = language or BY_NAME[name]["language_id"]
return _write(result, voice=name, language=label), _stats(result)
# --------------------------------------------------------------------------
# Clone. GPU. The microphone is the default path; an upload is gated.
# --------------------------------------------------------------------------
def _clone_duration(example, mic, upload, consent, text, language, seed, speed):
# Enrollment is a fixed cost on top of the render: two encoders and a
# tokenizer over at most 20 s of audio.
return _estimate(text, overhead=30.0)
@spaces.GPU(duration=_clone_duration)
def clone(example, mic, upload, consent: bool, text: str, language: str, seed, speed):
"""Enroll a voice, speak with it, keep nothing.
Three ways in, and they do not carry the same consent story, so they are
not collapsed into one input. A shipped example is a clip whose donor
released it for exactly this. A microphone recording is the visitor's own
voice, which is consent by construction. An upload is neither, so it is the
only one gated on a checkbox.
"""
if example:
# A file that ships in this repo. It must survive the request.
source = str(HERE / BY_NAME[example]["reference"]["public_preview"])
ephemeral = False
label = example
else:
source = mic or upload
ephemeral = True
label = "cloned"
if not source:
raise gr.Error(
"Pick an example, record yourself, or upload a clip you are allowed to use."
)
if upload and not mic and not consent:
raise gr.Error(
"Confirm the uploaded voice is yours, or that you have permission to use it."
)
text = _check(text, MAX_CLONE_CHARS)
if example and not language:
language = BY_NAME[example]["language_id"]
try:
import librosa
samples, _ = librosa.load(source, sr=24_000, mono=True)
limit = int(ENROLL_SECONDS * 24_000)
if samples.size > limit:
samples = samples[:limit]
try:
# The profile stays a local. It is never saved, never returned and
# never offered for download: the embeddings are the part of a
# cloned voice that would outlive the request if anything held them.
profile = enroller.enroll(samples, 24_000, name=label)
except ValueError as exc:
# The library's own messages name the bound and describe a good
# input, which is more useful than anything restated here.
raise gr.Error(str(exc)) from exc
# `enroll` writes no language, so every cloned voice would claim English
# and read its text through the English funnel.
profile = dataclasses.replace(profile, language=language or "en")
result = engine.synthesize_long(
text, profile, seed=int(seed), language=language or None, speed=float(speed)
)
return _write(result, voice=label, language=profile.language), _stats(result)
finally:
# Nothing the visitor recorded outlives the request that carried it.
# A shipped example is not the visitor's and is not ours to delete.
if ephemeral and source:
with contextlib.suppress(OSError):
os.unlink(source)
# --------------------------------------------------------------------------
# Determinism probe. GPU. Renders the same text twice at the same seed.
# --------------------------------------------------------------------------
def _probe_duration(text, name, seed):
return _estimate(text, passes=2, overhead=20.0)
@spaces.GPU(duration=_probe_duration)
def probe(text: str, name: str, seed: float):
text = _check(text, MAX_PROBE_CHARS)
profile = PROFILES[name]
first = engine.synthesize_long(text, profile, seed=int(seed))
second = engine.synthesize_long(text, profile, seed=int(seed))
left, right = _sha256_audio(first.audio), _sha256_audio(second.audio)
verdict = "Identical." if left == right else "Different. Please report this."
return "\n".join(
[
f"**{verdict}**",
"",
"```",
f"render 1 sha256 {left}",
f"render 2 sha256 {right}",
f" algo[{first.algorithm_fingerprint}] seed {int(seed)}",
"```",
"",
"Identical within this build and this device. loudkit promises a "
"bit-identical waveform for the same seed, build, backend and input. "
"It does not promise that your laptop matches this GPU. "
f"[Read the identity contract]({IDENTITY_CONTRACT}).",
]
)
# --------------------------------------------------------------------------
# Interface
# --------------------------------------------------------------------------
# loudreader.io: cream ground, ink text, black pill buttons at 14px.
CSS = """
#lk-head h1 { font-size: 2.15rem; margin-bottom: .25rem; letter-spacing: -.02em; }
#lk-head p { margin-top: 0; }
.lk-pill {
display: inline-block; padding: .2rem .75rem; margin: .15rem .35rem .15rem 0;
border: 1px solid #ded8ce; border-radius: 999px; font-size: .8rem;
color: #374151; background: #fffdfa;
}
.lk-card { background: #fffdfa; border: 1px solid #e7e1d7; border-radius: 14px; padding: .35rem 1rem; }
footer { display: none !important; }
"""
# Gradio follows the visitor's system theme unless told otherwise, and this
# palette is light-first. Without this the ink-on-cream tokens below land under
# a dark stylesheet and the text turns near-white on a cream ground.
FORCE_LIGHT = """
() => {
const url = new URL(window.location);
if (url.searchParams.get('__theme') !== 'light') {
url.searchParams.set('__theme', 'light');
window.location.replace(url.href);
}
}
"""
THEME = gr.themes.Soft(
primary_hue=gr.themes.colors.gray,
neutral_hue=gr.themes.colors.stone,
font=[gr.themes.GoogleFont("Inter"), "system-ui", "sans-serif"],
).set(
body_background_fill="#f7f5f2",
body_text_color="#111827",
body_text_color_subdued="#4b5563",
block_background_fill="#fffdfa",
block_border_color="#e7e1d7",
border_color_primary="#e7e1d7",
input_background_fill="#ffffff",
button_primary_background_fill="#111827",
button_primary_background_fill_hover="#374151",
button_primary_text_color="#ffffff",
button_large_radius="14px",
button_small_radius="14px",
)
with gr.Blocks(title="loudkit", theme=THEME, css=CSS, js=FORCE_LIGHT, fill_width=False) as demo:
gr.Markdown(
f"""
# Twenty voices. Ten languages. One engine.
On-device text to speech, running here on ZeroGPU.
[Model](https://huggingface.co/{REPO}) · [Code]({DOCS}) · [Responsible use]({RESPONSIBLE_USE})
Listening costs no GPU
Speaking and cloning spend your daily quota
algo[{FINGERPRINT}]
""",
elem_id="lk-head",
)
with gr.Tabs():
# ------------- Voices: listen for free, then type your own -------------
with gr.Tab("Voices"):
gr.Markdown(
"Pick a voice and the sample plays at once. Those are rendered "
"ahead of time and use no GPU. Type your own text underneath."
)
with gr.Row():
with gr.Column(scale=1):
with gr.Row():
lang_pick = gr.Dropdown(
LANGUAGE_FILTER, value=FIRST_LANGUAGE, label="Language"
)
pick = gr.Dropdown(
FIRST_VOICES, value=FIRST_VOICE, label="Voice"
)
made = gr.Audio(label="loudkit", type="filepath", interactive=False)
ref = gr.Audio(
label="Reference recording", type="filepath", interactive=False
)
with gr.Column(scale=1):
card = gr.Markdown(elem_classes="lk-card")
with gr.Accordion("The whole roster", open=False):
gr.Dataframe(
value=ROSTER_TABLE,
headers=["Voice", "Language", "Gender", "Licence", "Similarity"],
interactive=False,
wrap=True,
)
gr.Markdown("### Say something in this voice.")
gr.Markdown(
f"Up to {MAX_CHARS:,} characters here. The library itself takes 10,000. "
"This part spends your ZeroGPU quota."
)
say = gr.Textbox(
label="Your text",
placeholder="Hello from loudkit.",
lines=3,
# Without an explicit ceiling the box renders at its default
# maximum, which is twenty rows of empty space.
max_lines=6,
max_length=MAX_CHARS,
)
with gr.Row():
say_lang = gr.Dropdown(LANGUAGE_CHOICES, value="", label="Read the text as")
say_seed = gr.Number(value=7, precision=0, label="Seed")
say_speed = gr.Slider(
lk.MIN_SPEED, lk.MAX_SPEED, value=1.0, step=0.05, label="Speed"
)
say_go = gr.Button("Speak", variant="primary")
say_out = gr.Audio(label="Speech", type="filepath")
say_stats = gr.Markdown()
say_go.click(
speak, [say, pick, say_lang, say_seed, say_speed], [say_out, say_stats]
)
def on_language(language):
choices = voices_in(language)
name = choices[0][1]
return (gr.Dropdown(choices=choices, value=name), *listen(name))
lang_pick.change(on_language, lang_pick, [pick, made, ref, card])
pick.change(listen, pick, [made, ref, card])
demo.load(listen, pick, [made, ref, card])
with gr.Accordion("Determinism check", open=False):
gr.Markdown(
"This renders the same text twice at the same seed and hashes "
"both waveforms. The digests must match."
)
with gr.Row():
probe_text = gr.Textbox(
value="The same seed gives the same audio.",
label="Text",
lines=1,
max_lines=2,
max_length=MAX_PROBE_CHARS,
scale=3,
)
probe_seed = gr.Number(value=7, precision=0, label="Seed", scale=1)
probe_go = gr.Button("Render twice")
probe_out = gr.Markdown()
probe_go.click(probe, [probe_text, pick, probe_seed], probe_out)
# ---------------------------- Clone ----------------------------
with gr.Tab("Clone"):
gr.Markdown(
f"""
Clone a voice from a short recording, then speak with it.
- Try one of the shipped examples, or record yourself.
- Clone only your own voice, or a voice you have permission to use.
- Nothing is kept. The recording and the voice embeddings are discarded when the
request ends, and neither is offered for download.
- See [Responsible use]({RESPONSIBLE_USE}).
"""
)
with gr.Row():
with gr.Column(scale=1):
example = gr.Dropdown(
EXAMPLE_CHOICES,
value=CLONE_EXAMPLES[0],
label="Try an example",
info="Reference clips donated for building TTS voices.",
)
example_ref = gr.Audio(
label="What gets cloned",
type="filepath",
interactive=False,
show_download_button=False,
)
gr.Markdown("Or use your own voice. That clears the example.")
mic = gr.Audio(
sources=["microphone"], type="filepath", label="Record yourself"
)
with gr.Accordion("Upload a file instead", open=False):
upload = gr.Audio(
sources=["upload"], type="filepath", label="Audio file"
)
consent = gr.Checkbox(
value=False,
label=(
"This is my own voice, or I have permission from the "
"person who owns it."
),
)
with gr.Column(scale=1):
clone_text = gr.Textbox(
label="Text to speak",
placeholder="Now in my own voice.",
lines=3,
max_lines=6,
max_length=MAX_CLONE_CHARS,
)
clone_lang = gr.Dropdown(
LANGUAGE_CHOICES, value="", label="Language of the text"
)
with gr.Row():
clone_seed = gr.Number(value=7, precision=0, label="Seed")
clone_speed = gr.Slider(
lk.MIN_SPEED, lk.MAX_SPEED, value=1.0, step=0.05, label="Speed"
)
clone_go = gr.Button("Clone and speak", variant="primary")
clone_out = gr.Audio(
label="Speech", type="filepath", show_download_button=False
)
clone_stats = gr.Markdown()
def show_example(name):
if not name:
return None
return str(HERE / BY_NAME[name]["reference"]["public_preview"])
def clear_example(value):
# Recording or uploading takes over from the example, so the two
# cannot both be armed and leave the visitor guessing which won.
return gr.Dropdown(value="") if value else gr.skip()
example.change(show_example, example, example_ref)
demo.load(show_example, example, example_ref)
mic.change(clear_example, mic, example)
upload.change(clear_example, upload, example)
clone_go.click(
clone,
[example, mic, upload, consent, clone_text, clone_lang, clone_seed, clone_speed],
[clone_out, clone_stats],
)
gr.Markdown(
f"""
---
Run the same engine locally, where nothing is queued and nothing is metered.
```bash
pip install "loudkit[torch,audio,enroll,hub]"
```
```python
import loudkit as lk
engine = lk.load("{REPO}")
voice = lk.voice("kathleen", repo="{REPO}")
engine.synthesize_long("Hello from loudkit.", voice, seed=7).save("hello.wav")
```
Output files carry C2PA provenance: the fingerprint, the recipe and the seed.
"""
)
# The engine holds one set of weights and renders with an internal producer
# thread. One render at a time keeps two requests off the same buffers.
demo.queue(default_concurrency_limit=1, max_size=24)
if __name__ == "__main__":
demo.launch()