wfloat

wfloat is the Python package for Wfloat's on-device model families, including TTS, STT, VAD, and LLM.

It runs inference locally in Python instead of calling a hosted inference API. The TTS model supports 20 voices with emotion and intensity control.

If you're building for the browser, use @wfloat/wfloat-web. If you're building for React Native, use @wfloat/react-native-wfloat.

Browser demo to hear how it sounds: https://wfloat.com/demo

Install

pip install wfloat

Usage

import wfloat

tts = wfloat.load_tts_model("wfloat/wfloat-tts")

result = tts.synthesize(
    text="No, no, that's not possible.",
    voice="mad_scientist_woman",
    emotion="surprise",
    intensity=0.7,
)

print(result.model_id)
print(result.timeline.chunks[0].text)

For multi-speaker dialogue:

import wfloat

tts = wfloat.load_tts_model("wfloat/wfloat-tts")

result = tts.synthesize_dialogue(
    segments=[
        {
            "voice": "wise_elder_man",
            "text": "Rain taps against the tavern shutters as you step inside.",
            "emotion": "neutral",
            "intensity": 0.5,
        },
        {
            "voice": "strong_hero_man",
            "text": "You're late. Two bandits stole the king's map over three hours ago.",
            "emotion": "fear",
            "intensity": 0.6,
        },
        {
            "voice": "strong_hero_man",
            "text": "They fled north, up into the woods.",
            "emotion": "neutral",
            "intensity": 0.5,
        },
    ],
    silence_between_segments_sec=0.35,
)

result.audio.save("dialogue.wav")

The older load(...), generate(...), and generate_dialogue(...) names are still available as compatibility aliases.

STT Usage

The shared Python entrypoint is load_stt_model(...), with load_whisper_tiny_en(...) as a convenience wrapper for offline STT:

import wfloat

stt = wfloat.load_stt_model(
    "openai/whisper-tiny-en",
)

result = stt.transcribe(audio="/path/to/audio.wav")
print(result.text)

The loader accepts canonical built-in model IDs and resolves Wfloat-hosted registry assets internally.

Streaming-capable STT families also expose a separate session path instead of overloading transcribe(...):

stt = wfloat.load_stt_model("k2-fsa/streaming-zipformer-en")
session = stt.create_session()

session.push(audio_chunk, sample_rate=16000)
partial = session.get_result()
final_result = session.finish()
session.close()

VAD Usage

Python also exposes the same one-shot VAD model shape as the web and React Native packages. It is intentionally file/buffer based for now; there is no Python live microphone/session helper. The Python VAD path uses wfloat-core, matching the TTS, STT, and LLM backend boundary.

vad = wfloat.load_vad_model(
    "snakers4/silero-vad",
    threshold=0.5,
    min_silence_duration_sec=0.5,
    min_speech_duration_sec=0.25,
    max_speech_duration_sec=20.0,
)

result = vad.detect(audio="/path/to/mono-16khz.wav")

for segment in result.segments:
    print(segment.start_sec, segment.duration_sec)

VAD currently expects mono 16 kHz audio.

LLM Usage

The Python LLM path loads local GGUF artifacts through wfloat-core:

import wfloat

llm = wfloat.load_llm_model("HuggingFaceTB/SmolLM2-360M-Instruct")
result = llm.generate("Write one calm sentence about local inference.", seed=0)
print(result.text)
llm.close()

CLI Usage

You can also generate a WAV from the command line:

wfloat generate \
  --text "Hello world!" \
  --out out.wav \
  --voice-id mad_scientist_woman \
  --emotion surprise \
  --intensity 0.7 \
  --silence-padding-sec 0

For the full CLI help:

wfloat generate --help

The first load downloads the model assets. After that, the package uses the cached local copy.

Native Backend

Python TTS, STT, VAD, and LLM use the wfloat-core native runtime. Release wheels bundle the platform-specific shared library inside the wfloat package.

Inside this monorepo, local development can point at an explicit build artifact with:

export WFLOAT_CORE_LIBRARY=/abs/path/to/libwfloat-core.so

Speaker IDs

Use voice_id string names or numeric sid values:

Speaker	SID
skilled_hero_man	0
skilled_hero_woman	1
fun_hero_man	2
fun_hero_woman	3
strong_hero_man	4
strong_hero_woman	5
mad_scientist_man	6
mad_scientist_woman	7
clever_villain_man	8
clever_villain_woman	9
narrator_man	10
narrator_woman	11
wise_elder_man	12
wise_elder_woman	13
outgoing_anime_man	14
outgoing_anime_woman	15
scary_villain_man	16
scary_villain_woman	17
news_reporter_man	18
news_reporter_woman	19

Emotions

Supported emotion labels:

• neutral
• joy
• sadness
• anger
• fear
• surprise
• dismissive
• confusion

intensity must be between 0.0 and 1.0.

More

• Docs: https://docs.wfloat.com
• Model card, voices, emotions, and samples: https://huggingface.co/Wfloat/wfloat-tts
• Web package: https://github.com/wfloat/wfloat-web
• React Native package: https://github.com/wfloat/react-native-wfloat