KugelAudio Python SDK

Official Python SDK for the KugelAudio Text-to-Speech API.

Installation

pip install kugelaudio

Or with uv:

uv add kugelaudio

Quick Start

from kugelaudio import KugelAudio

# Initialize the client - just needs an API key!
client = KugelAudio(api_key="your_api_key")

# Generate speech
audio = client.tts.generate(
    text="Hello, world!",
    model_id="kugel-3",
    voice_id=1071,     # browse voices with client.voices.list()
    language="en",     # skip auto-detection (~150ms) when you know the language
)

# Save to file
audio.save("output.wav")

voice_id is required — 1071 ("Samantha Ferris", en-US) is a public voice you can call straight away. List every voice available to your key with client.voices.list() — see Voices.

Local CPU Turn Detection

The optional turn-detection runtime downloads the private, version-pinned ONNX bundle from Hugging Face, verifies every declared SHA-256 checksum, and then runs fully locally on CPU. ONNX Runtime performs model inference; the extra also loads the CPU PyTorch runtime to reproduce the native numerical environment used for the published quality gates. It requires Python 3.11 or newer.

pip install "kugelaudio[turn-detection]"
hf auth login  # required while the model repository is private

Load one model per process and create cheap state for each conversation:

from kugelaudio.turn import TurnDecisionReason, TurnDetector, TurnOutcomeKind

detector = TurnDetector.from_pretrained(cpu_threads=12)
turn = detector.create_session("en")

# Feed the current user's audio and interim ASR transcript continuously.
turn.push_pcm16(pcm_chunk, sample_rate=16_000)
turn.update_transcript("I think we should probably")

# Call with the current duration whenever VAD reports real silence.
decision = turn.observe_silence(duration_ms=200)

if not decision.end_turn:
    # Temporal bundles rescore ambiguous evidence at their later policy points.
    decision = turn.observe_silence(duration_ms=300)

if decision.end_turn:
    start_assistant_response()
    turn.reset_turn()

Stable v2.1.1 contains both qualified quantization variants on one immutable policy-only patch over the byte-identical v2.1.0 weights. W8A32 is recommended and selected by default:

detector = TurnDetector.from_pretrained(
    revision="v2.1.1",
    preset="responsive-300ms",  # or conservative-600ms
    cpu_threads=12,
)

The supported W8A8 QAT variant must be selected explicitly because W8A32 has better endpoint quality:

quantized = TurnDetector.from_pretrained(
    revision="v2.1.1",
    variant="models/w8a8-qat",
    preset="responsive-300ms",
    cpu_threads=12,
)

The SDK default pins the exact commit behind stable tag v2.1.1. Human-facing release tags are immutable; channels/stable and channels/preview are mutable operational pointers and must be selected explicitly. The W8A8 model is stable supported but not recommended: its 11.21% responsive false-cutoff rate exceeds the 9% gate, while W8A32 reaches 8.65%. Unqualified models live under experiments/* and are never selected by the default runtime.

With correctly confirmed false-cutoff feedback, the W8A32 policy measures 7.94% at 296.5 ms and 3.55% at 594.5 ms on the same public validation grid. These are conditional results selected on public validation, not an automatic feedback classifier or an independent production-generalization claim.

On Linux x86_64 the turn-detection extra installs the calibrated OpenVINO runtime; other platforms retain the CPU reference runtime. A policy requiring OpenVINO fails explicitly when that provider is unavailable. For native macOS development, opt into the verified but quality-unqualified CPU path explicitly:

detector = TurnDetector.from_pretrained(execution_provider="cpu")

Omitting the override always retains the bundle-calibrated provider. The qualified runtime uses BF16 activations only for the fixed-shape Whisper stage; the dynamic-text fusion stage stays FP32 to avoid non-finite OpenVINO outputs. Servers with a compatible onnxruntime-gpu installation may explicitly select execution_provider="cuda". CUDA selection fails during model loading if the provider is unavailable. The qualified graphs keep their expensive encoder and matrix kernels on CUDA while ONNX Runtime handles small dynamic-shape/control nodes on an explicit CPU fallback; the standard turn-detection extra remains the qualified local CPU/OpenVINO distribution.

cpu_streams=1 gives the lowest single-conversation latency. Shared workers can set cpu_streams=2 for two concurrent turns or cpu_streams=4 for four or more; streams improve parallel request throughput but increase isolated-call latency because the fixed CPU thread budget is divided between them. On the qualified 26-core Xeon target, cpu_threads=20 is the balanced optimum: use it only when the worker owns at least 20 physical cores, and keep the explicit lower setting for smaller allocations.

Temporal sessions reuse exact intermediate work without weakening rescoring. If neither audio nor transcript changed, the next duration-specific threshold reuses the identical probability; a transcript-only revision reruns Qwen while reusing Whisper embeddings; appended audio always reruns both graphs. On the qualified 20-core target these paths measured 0.0 ms, 20.4 ms, and 39.9 ms p50, respectively, after a 45.3 ms first score.

For a latency-first application, commit a confident completion immediately when the first 200 ms score returns while retaining the calibrated threshold and incomplete-turn timeout:

turn = detector.create_session("en", action_delay_ms=200)

The incomplete-turn fallback defaults to the bundle's calibrated timeout. Pass timeout_ms only when the surrounding product has a measured override. Read the resolved value from effective_timeout_ms on TurnSession, KugelTurnBridge, or KugelTurnStopStrategy.

When VAD detects speech before the endpoint action, cancel the pending decision:

turn.speech_resumed()

Temporal policies automatically learn a speaker's ordinary within-turn pause lengths when speech resumes. If product-level evidence distinguishes a likely false cutoff from an intentional barge-in, record that outcome explicitly:

turn.record_outcome(
    TurnOutcomeKind.LIKELY_FALSE_CUTOFF,
    speaker_pause_ms=420,
)

Do not infer false-cutoff feedback from speech resumption alone; intentional barge-ins must be recorded as CONFIRMED_BARGE_IN or left neutral.

Important input and lifecycle rules:

• Audio must be mono 16 kHz. push_audio accepts normalized float32; use push_pcm16 for explicit little-endian signed PCM16 conversion.
• The transcript must be the current interim user transcript. Do not wait for a post-endpoint final transcript.
• Legacy policies score each silence episode once. Temporal policies rescore at each declared silence duration (for example 200/300/400/600 ms), using the latest audio/transcript snapshot and stopping after the first confident score.
• Call observe_silence with monotonically increasing durations for one silence span. Call speech_resumed before starting a new span.
• Only P(complete) can trigger a model endpoint. incomplete, backchannel, and wait keep listening until speech resumes or the measured timeout fires.
• One TurnDetector owns the roughly 1.55 GiB model runtime. Share it across sessions instead of loading one copy per conversation.
• No network is used after the immutable model revision is cached. Pass local_files_only=True to enforce offline startup.

The stable v2.1.1 policies cover English. Applications serving additional languages must explicitly route those sessions to a calibrated multilingual revision; the SDK never silently changes model identity. Unsupported languages, missing private-repository access, corrupt bundles, wrong sample rates, and invalid session ordering raise typed TurnDetectionError subclasses.

LiveKit Agents

Install both optional integrations and LiveKit's VAD plugin:

pip install "kugelaudio[livekit,turn-detection]" "livekit-agents[silero]"

KugelTurnBridge keeps LiveKit's normal STT pipeline intact while teeing its audio into KugelTurn. Configure LiveKit for manual endpointing so two detectors cannot commit the same user turn:

import logging
from collections.abc import AsyncIterable

from livekit import rtc
from livekit.agents import Agent, AgentSession, ModelSettings
from livekit.plugins import silero

from kugelaudio.livekit import KugelTurnBridge
from kugelaudio.turn import TurnDetector

logger = logging.getLogger(__name__)
detector = TurnDetector.from_pretrained(cpu_threads=12)  # once per worker


def observe_decision(decision):
    logger.info(
        "turn decision",
        extra={
            "reason": decision.reason.value,
            "end_turn": decision.end_turn,
            "silence_ms": decision.silence_ms,
            "inference_ms": decision.inference_ms,
            "probabilities": decision.probabilities,
        },
    )


bridge = KugelTurnBridge(
    detector,
    language="en",
    vad_silence_ms=200,
    action_delay_ms=200,  # omit to use the calibrated language delay
    on_decision=observe_decision,
)


class VoiceAgent(Agent):
    def stt_node(
        self,
        audio: AsyncIterable[rtc.AudioFrame],
        model_settings: ModelSettings,
    ):
        return bridge.stt_node(self, audio, model_settings)


session = AgentSession(
    turn_detection="manual",
    vad=silero.VAD.load(min_silence_duration=0.2),
    stt=stt,
    llm=llm,
    tts=tts,
)

try:
    await session.start(agent=VoiceAgent(instructions="..."), room=ctx.room)
finally:
    await bridge.aclose()

On LiveKit Agents 1.5+, the equivalent non-deprecated configuration is turn_handling=TurnHandlingOptions(turn_detection="manual"). The bridge resamples mono LiveKit input to 16 kHz, accumulates interim/final transcripts, runs ONNX inference outside the event loop, cancels pending decisions when speech resumes, and calls commit_user_turn() when the measured policy ends.

Keep vad_silence_ms equal to LiveKit's min_silence_duration in milliseconds. The validated configuration is 200 ms.

on_decision runs for the first model score and each meaningful policy-state change, including the final model_complete or timeout. Its TurnDecision contains the transcript, threshold, four class probabilities, silence duration, and inference_ms when a new model score was computed. Keep the callback non-blocking; enqueue network or storage work instead of performing it inline.

Pipecat

Install the Pipecat and turn-detection extras:

pip install "kugelaudio[pipecat,turn-detection]"

Use KugelTurnStopStrategy as the sole Pipecat user-turn stop strategy:

from pipecat.audio.vad.silero import SileroVADAnalyzer
from pipecat.audio.vad.vad_analyzer import VADParams
from pipecat.processors.aggregators.llm_response_universal import (
    LLMContextAggregatorPair,
    LLMUserAggregatorParams,
)
from pipecat.turns.user_turn_strategies import UserTurnStrategies

from kugelaudio.pipecat import KugelTurnStopStrategy
from kugelaudio.turn import TurnDetector

detector = TurnDetector.from_pretrained(cpu_threads=12)  # once per process
turn_strategy = KugelTurnStopStrategy(
    detector,
    language="en",
    vad_silence_ms=200,
    action_delay_ms=200,  # omit to use the calibrated language delay
)

user_aggregator, assistant_aggregator = LLMContextAggregatorPair(
    context,
    user_params=LLMUserAggregatorParams(
        vad_analyzer=SileroVADAnalyzer(
            params=VADParams(stop_secs=0.2),
        ),
        user_turn_strategies=UserTurnStrategies(
            stop=[turn_strategy],
        ),
    ),
)

The strategy supports Pipecat 0.0.101+ and 1.x turn-management APIs. It consumes native InputAudioRawFrame, transcription, and VAD frames, explicitly rejects multichannel input, resamples mono PCM to 16 kHz, and emits Pipecat's standard on_user_turn_stopped event. For Pipecat 0.x releases whose VAD stop frame does not carry its own duration, keep vad_silence_ms equal to VADParams.stop_secs.

Client Configuration

from kugelaudio import KugelAudio

# Simple setup - single URL handles everything
client = KugelAudio(api_key="your_api_key")

# Or with custom options
client = KugelAudio(
    api_key="your_api_key",           # Required: Your API key
    api_url="https://api.kugelaudio.com",  # Optional: API base URL (default)
    timeout=60.0,                      # Optional: Request timeout in seconds
)

Region Selection

By default, KugelAudio uses the canonical geo-routed API endpoint. You can select the direct EU endpoint when you need to pin traffic to Europe.

Option 1 — API key prefix (simplest, works with env vars):

client = KugelAudio(api_key="eu-ka_your_api_key")       # → EU
client = KugelAudio(api_key="ka_your_api_key")          # → canonical geo-routed API

Option 2 — region parameter:

client = KugelAudio(api_key="ka_your_api_key", region="eu")

The prefix is always stripped before authentication. Priority: api_url > region > key prefix > default.

Single URL Architecture

The SDK uses a single URL for both REST API and WebSocket streaming. The TTS server provides both REST endpoints (/v1/models, /v1/voices) and WebSocket (/ws/tts) - no proxy needed, minimal latency.

Local Development

For local development, point directly to your TTS server:

client = KugelAudio(
    api_key="your_api_key",
    api_url="http://localhost:8000",   # TTS server handles everything
)

Or if you have separate backend and TTS servers:

client = KugelAudio(
    api_key="your_api_key",
    api_url="http://localhost:8001",   # Backend for REST API
    tts_url="http://localhost:8000",   # TTS server for WebSocket streaming
)

Available Models

kugel-3 is the current production model — use it for new integrations. Legacy IDs (kugel-2.5, kugel-2-turbo, kugel-2, kugel-1, kugel-1-turbo) are still accepted for backwards compatibility; see Models.

List Available Models

models = client.models.list()

for model in models:
    print(f"{model.id}: {model.name}")
    print(f"  Description: {model.description}")
    print(f"  Max Input: {model.max_input_length} characters")
    print(f"  Sample Rate: {model.sample_rate} Hz")

Voices

List Available Voices

# List all available voices (paginated)
result = client.voices.list()

for voice in result.voices:
    print(f"{voice.id}: {voice.name}")
    print(f"  Category: {voice.category}")
    print(f"  Languages: {', '.join(voice.supported_languages)}")
print(f"Showing {len(result.voices)} of {result.total} voices")

# Filter by language
result = client.voices.list(language="de")

# Get only public voices
result = client.voices.list(include_public=True)

# Paginate through results
page1 = client.voices.list(limit=10, offset=0)
page2 = client.voices.list(limit=10, offset=10)

Get a Specific Voice

voice = client.voices.get(voice_id=1071)
print(f"Voice: {voice.name}")
print(f"Sample text: {voice.sample_text}")

Text-to-Speech Generation

Basic Generation (Non-Streaming)

Generate complete audio and receive it all at once:

audio = client.tts.generate(
    text="Hello, this is a test of the KugelAudio text-to-speech system.",
    model_id="kugel-3",        # Current model (see "Available Models")
    voice_id=1071,             # Required: voice to speak with (client.voices.list())
    cfg_scale=2.0,             # Guidance scale (1.2-2.5)
    max_new_tokens=2048,       # Maximum tokens to generate
    sample_rate=24000,         # Output sample rate
    output_format=None,        # Optional: 'pcm_24000', 'ulaw_8000', 'alaw_8000', ...
    normalize=True,            # Enable text normalization (see below)
    language="en",             # Language for normalization
)

# Audio properties
print(f"Duration: {audio.duration_seconds:.2f}s")
print(f"Samples: {audio.samples}")
print(f"Sample rate: {audio.sample_rate} Hz")
print(f"Generation time: {audio.generation_ms:.0f}ms")
print(f"RTF: {audio.rtf:.2f}")  # Real-time factor

# Save to WAV file
audio.save("output.wav")

# Get raw PCM bytes
pcm_data = audio.audio

# Get WAV bytes (with header)
wav_bytes = audio.to_wav_bytes()

Streaming Audio Output

Receive audio chunks as they are generated for lower latency:

# Synchronous streaming
for item in client.tts.stream(
    text="Hello, this is streaming audio.",
    model_id="kugel-3",
    voice_id=1071,
    language="en",
):
    if hasattr(item, 'audio'):  # AudioChunk
        # Process audio chunk immediately
        print(f"Chunk {item.index}: {len(item.audio)} bytes, {item.samples} samples")
        # play_audio(item.audio)
    elif isinstance(item, dict) and item.get('final'):
        # Final stats
        print(f"Total duration: {item.get('dur_ms', 0):.0f}ms")
        print(f"Generation time: {item.get('gen_ms', 0):.0f}ms")

Async Streaming

For async applications:

import asyncio

async def generate_speech():
    async for item in client.tts.stream_async(
        text="Async streaming example.",
        model_id="kugel-3",
        voice_id=1071,
        language="en",
    ):
        if hasattr(item, 'audio'):
            # Process chunk
            pass

asyncio.run(generate_speech())

Async Generation

import asyncio

async def main():
    audio = await client.tts.generate_async(
        text="Async generation example.",
        model_id="kugel-3",
        voice_id=1071,
        language="en",
    )
    audio.save("async_output.wav")

asyncio.run(main())

Text Normalization

Text normalization converts numbers, dates, times, and other non-verbal text into spoken words. For example:

• "I have 3 apples" → "I have three apples"
• "The meeting is at 2:30 PM" → "The meeting is at two thirty PM"
• "€50.99" → "fifty euros and ninety-nine cents"

Usage

# With explicit language (recommended - fastest)
audio = client.tts.generate(
    text="I bought 3 items for €50.99 on 01/15/2024.",
    voice_id=1071,
    normalize=True,
    language="en",  # Specify language for best performance
)

# With auto-detection (adds ~150ms latency)
audio = client.tts.generate(
    text="Ich habe 3 Artikel für 50,99€ gekauft.",
    voice_id=1705,  # a German voice — client.voices.list(language="de")
    normalize=True,
    # language not specified - will auto-detect
)

Supported Languages

Performance Warning

⚠️ Latency Warning: Using normalize=True without specifying language adds approximately 150ms latency for language auto-detection. For best performance in latency-sensitive applications, always specify the language parameter.

LLM Integration: Streaming Text Input

For real-time TTS when streaming text from an LLM (GPT-4, Claude, etc.), use a StreamingSession. Forward LLM tokens directly to session.send() without flush=True — the server accumulates them and starts generation at natural sentence boundaries. Flush exactly once at the end of the assistant turn.

⚠️ Do not call session.send(text, flush=True) between sentences or words. Each explicit flush is a separate TTS request that pays the full model time-to-first-audio (TTFA) again and produces an audible gap. See Chunking & per-segment latency for the full rationale and ElevenLabs migration notes.

Async Streaming Session

import asyncio

async def speak_turn(llm_token_stream):
    async with client.tts.streaming_session(
        voice_id=1071,
        model_id="kugel-3",
        language="en",
    ) as session:
        # Forward every LLM token directly. No flush=True per token —
        # the server's text buffer chunks at sentence boundaries.
        async for token in llm_token_stream:
            async for chunk in session.send(token):
                play_audio(chunk.audio)

        # Single flush at turn end emits any trailing text.
        async for chunk in session.flush():
            play_audio(chunk.audio)

        # Per-session usage — bill your own customers per conversation.
        # cost_cents is the actual charge in EUR cents (None if undetermined).
        usage = session.last_usage
        if usage:
            print(f"audio: {usage.audio_seconds}s, cost: {usage.cost_cents} ct")

asyncio.run(speak_turn(my_llm_stream()))

Synchronous Streaming Session

with client.tts.streaming_session_sync(
    voice_id=1071,
    model_id="kugel-3",
    language="en",
) as session:
    for token in llm_token_stream:
        for chunk in session.send(token):  # no flush per token
            play_audio(chunk.audio)

    for chunk in session.flush():  # single flush at turn end
        play_audio(chunk.audio)

Pipecat Word Timestamps

The Pipecat service can forward server word timings with their KugelAudio context ID. Supplying the callback enables timestamp generation; omitting it keeps the existing audio-only behavior.

from kugelaudio.models import WordTimestamp
from kugelaudio.pipecat import KugelAudioTTSService

def on_word_timestamps(
    context_id: str, timestamps: list[WordTimestamp]
) -> None:
    for timestamp in timestamps:
        record_timing(context_id, timestamp)

tts = KugelAudioTTSService(
    api_key="your_api_key",
    voice_id=123,
    on_word_timestamps=on_word_timestamps,
)

Error Handling

from kugelaudio import KugelAudio
from kugelaudio.exceptions import (
    KugelAudioError,
    AuthenticationError,
    RateLimitError,
    InsufficientCreditsError,
    ValidationError,
    NotFoundError,
)

try:
    audio = client.tts.generate(text="Hello!", voice_id=1071)
except AuthenticationError:
    print("Invalid API key")
except RateLimitError:
    print("Rate limit exceeded, please wait")
except InsufficientCreditsError:
    print("Not enough credits, please top up")
except ValidationError as e:
    print(f"Invalid request: {e}")
except NotFoundError as e:
    print(f"Resource not found (e.g. unknown voice_id): {e}")
except KugelAudioError as e:
    print(f"API error: {e}")

Rolling deploys

When a replica restarts, the server closes the WebSocket with code 1012 (or refuses the upgrade with 1013). StreamingSession and MultiContextSession handle that for you: if no audio for the current turn has been delivered yet, they wait one second, reconnect to another replica and resend the turn, session config and open contexts included, then keep streaming. The replay is logged at INFO on the kugelaudio.streaming logger and your code sees nothing else.

A turn whose audio already started is not replayed, because that would repeat what the listener just heard, and no turn is replayed twice. Those two cases raise ServerRestartingError, a ConnectionError with retry_after = 1:

from kugelaudio.exceptions import ServerRestartingError

try:
    async for chunk in session.send("Hello!", flush=True):
        play(chunk)
except ServerRestartingError as e:
    await asyncio.sleep(e.retry_after or 1)
    await session.connect()  # a fresh replica; resend what was not spoken

Data Models

AudioChunk

Represents a single audio chunk from streaming:

class AudioChunk:
    audio: bytes          # Raw PCM16 audio data
    encoding: str         # 'pcm_s16le' | 'mulaw' | 'alaw' (G.711 when output_format set)
    index: int           # Chunk index (0-based)
    sample_rate: int     # Sample rate (24000)
    samples: int         # Number of samples in chunk

    @property
    def duration_seconds(self) -> float:
        """Duration of this chunk in seconds."""

AudioResponse

Complete audio response from generation:

class AudioResponse:
    audio: bytes              # Complete PCM16 audio
    sample_rate: int          # Sample rate (24000)
    samples: int              # Total samples
    duration_ms: float        # Duration in milliseconds
    generation_ms: float      # Generation time in milliseconds
    rtf: float               # Real-time factor

    @property
    def duration_seconds(self) -> float:
        """Duration in seconds."""

    def save(self, path: str) -> None:
        """Save as WAV file."""

    def to_wav_bytes(self) -> bytes:
        """Get WAV file as bytes."""

Model

TTS model information:

class Model:
    id: str                   # 'kugel-3' (or a legacy ID such as 'kugel-2.5')
    name: str                 # Human-readable name
    description: str          # Model description
    max_input_length: int     # Maximum input characters
    sample_rate: int          # Output sample rate

Supported native output format tokens are pcm_8000, pcm_16000, pcm_22050, pcm_24000, ulaw_8000, and alaw_8000.

Voice

Voice information:

class Voice:
    id: int                          # Voice ID
    name: str                        # Voice name
    description: Optional[str]       # Description
    category: Optional[VoiceCategory]  # 'premade', 'cloned', 'generated'
    sex: Optional[VoiceSex]          # 'male', 'female', 'neutral'
    age: Optional[VoiceAge]          # 'young', 'middle_aged', 'old'
    supported_languages: List[str]   # ['en', 'de', ...]
    sample_text: Optional[str]       # Sample text for preview
    avatar_url: Optional[str]        # Avatar image URL
    sample_url: Optional[str]        # Sample audio URL
    is_public: bool                  # Whether voice is public
    verified: bool                   # Whether voice is verified

Complete Example

from kugelaudio import KugelAudio

# Initialize client
client = KugelAudio(api_key="your_api_key")

# List available models
print("Available Models:")
for model in client.models.list():
    print(f"  - {model.id}: {model.name}")

# List available voices
print("\nAvailable Voices:")
for voice in client.voices.list(limit=5).voices:
    print(f"  - {voice.id}: {voice.name}")

# Generate audio
print("\nGenerating audio...")
audio = client.tts.generate(
    text="Welcome to KugelAudio. This is an example of high-quality text-to-speech synthesis.",
    model_id="kugel-3",
    voice_id=1071,
    language="en",
)

print(f"Generated {audio.duration_seconds:.2f}s of audio in {audio.generation_ms:.0f}ms")
print(f"Real-time factor: {audio.rtf:.2f}x")

# Save to file
audio.save("example.wav")
print("Saved to example.wav")

# Close client
client.close()

Verifying the AI-generated watermark

Audio produced by the API carries an in-band watermark (EU AI Act Art. 50) alongside the X-KugelAudio-AI-Generated: true response header. The optional watermark extra verifies a clip locally.

pip install "kugelaudio[watermark]"

from kugelaudio.watermark import WatermarkDetector

detector = WatermarkDetector()          # load once, reuse
result = detector.detect_file("speech.wav")

if result.ai_generated:
    print(f"AI-generated (confidence {result.confidence:.3f})")
    print(f"source id {result.customer_id}")

detect(samples, sample_rate) scores an in-memory mono float array at any rate. It resamples to the watermark's native 24 kHz before scoring. There is one bundled KugelAudio detector; it is numpy-only, works offline, and does not take a backend argument.

The payload is a 12-bit source id (0–4095) encoded as a codeword across 30 detector windows (roughly four seconds), rather than as two raw bits. Presence can be reported from a shorter clip, but customer_id remains unset until there is enough aligned payload evidence to attribute the source.

Scoring is duration-dependent — more audio means more windows to average — so clips shorter than one second are rejected rather than answered unreliably; pass min_seconds= if you accept a weaker verdict. Multi-channel, integer, empty, non-finite, and out-of-range audio raises WatermarkAudioError instead of being silently coerced, because each of those would change the verdict you are about to act on.

Agent skill

If you build with a coding agent, install the bundled skill so it gets the TTFA rules, streaming semantics, and text-formatting constraints without you re-explaining them:

kugelaudio-skills install     # → ./.claude/skills/kugelaudio-tts/

--global installs to ~/.claude/skills/, and --dest <dir> selects another target.

License

MIT