babelvox 0.6.1

Real-time text-to-speech on Intel NPU/CPU via OpenVINO

BabelVox

Real-time text-to-speech on Intel NPU via OpenVINO. Runs Qwen3-TTS 0.6B inference entirely on Intel NPU (AI Boost), achieving RTF=1.0x (real-time) speech synthesis on a Lunar Lake ultrabook.

No PyTorch at runtime. Dependencies: openvino, numpy, librosa, soundfile, scipy, transformers (tokenizer only).

Installation

pip install babelvox

Or from source:

git clone https://github.com/Djwarf/babelvox.git
cd babelvox
pip install -e .

Quick start

Models (~2.5 GB) are downloaded automatically from HuggingFace on first run and cached for future use. No manual setup needed.

As a library

from babelvox import BabelVox
import soundfile as sf

# CPU — works on any machine (models auto-download on first use)
tts = BabelVox(precision="int8", use_cp_kv_cache=True)
wav, sr = tts.generate("Don't panic.", language="English")
sf.write("output.wav", wav, sr)

For Intel NPU (Lunar Lake or later), enable hardware acceleration and model caching:

tts = BabelVox(device="NPU", precision="int8",
               use_cp_kv_cache=True, talker_buckets=[64, 128, 256],
               cache_dir="./ov_cache")

From the command line

# CPU (works anywhere, ~1.1x RTF)
babelvox --int8 --cp-kv-cache --text "Hello world" --output hello.wav

# Intel NPU (real-time, RTF=1.0x)
babelvox --device NPU --int8 --cp-kv-cache --talker-buckets "64,128,256" \
  --cache-dir ./ov_cache \
  --text "Hello, this is real-time speech synthesis on an Intel NPU." \
  --output hello.wav

Features

Voice cloning

Clone any voice from a short reference audio clip (3-10 seconds). For best results, provide a transcription of the reference audio with ref_text:

wav, sr = tts.generate(
    "This sounds like someone else.",
    ref_audio="reference.wav",
    ref_text="The words spoken in reference dot wav.",
    language="English",
)

babelvox --int8 --cp-kv-cache --ref-audio reference.wav \
  --ref-text "The words spoken in reference dot wav." \
  --text "This sounds like someone else." --output cloned.wav

Voice persistence

Without a reference audio, each generate() call may produce a different voice. To keep a consistent voice across multiple calls:

# Extract once from reference audio, reuse for all subsequent calls
tts.default_speaker = tts.extract_speaker_embedding("voice.wav")
wav1, sr = tts.generate("First sentence.")
wav2, sr = tts.generate("Second sentence.")  # same voice

# Or pass a speaker embedding directly per call
embed = tts.extract_speaker_embedding("voice.wav")
wav, sr = tts.generate("Hello", speaker_embed=embed)

Model caching

On NPU, OpenVINO compiles models at startup which can take minutes on first run. Use cache_dir to cache compiled models so subsequent launches are instant:

tts = BabelVox(device="NPU", cache_dir="./ov_cache", precision="int8")
# First run: ~200s compile. Second run: instant.

babelvox --device NPU --cache-dir ./ov_cache --int8 --cp-kv-cache

Streaming generation

Start playback immediately instead of waiting for the full utterance. generate_stream() yields waveform chunks every N codec frames (~1 second per 12 frames):

import sounddevice as sd

for chunk, sr in tts.generate_stream("A long paragraph of text...",
                                      chunk_frames=12):
    sd.play(chunk, sr)
    sd.wait()

Silence-aware chunking

Split at natural pauses between words instead of fixed intervals. Produces complete words/phrases per chunk with crossfade overlap for click-free playback:

for chunk, sr in tts.generate_stream(
    "A long paragraph of text...",
    split_on_silence=True,   # yield at pauses between words
    min_chunk_frames=6,      # at least 0.5s per chunk
    max_chunk_frames=48,     # at most 4s per chunk
    silence_threshold=0.02,  # RMS energy threshold for silence
    crossfade_samples=1200,  # 50ms overlap for click-free joins
):
    sd.play(chunk, sr)
    sd.wait()

For near-zero gap between chunks, overlap generation and playback with a thread:

import threading, queue

audio_q = queue.Queue()

def producer():
    for chunk, sr in tts.generate_stream("Long text...",
                                          split_on_silence=True):
        audio_q.put((chunk, sr))
    audio_q.put(None)

threading.Thread(target=producer).start()

while True:
    item = audio_q.get()
    if item is None:
        break
    sd.play(item[0], item[1])
    sd.wait()

10 languages

Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish.

wav, sr = tts.generate("Bonjour le monde.", language="French")
wav, sr = tts.generate("Hallo Welt.", language="German")

Sampling controls

Fine-tune the generation quality and diversity:

wav, sr = tts.generate(
    "Hello world",
    temperature=0.9,          # higher = more expressive, lower = more stable
    top_k=50,                 # limit sampling to top-k tokens
    top_p=1.0,                # nucleus sampling threshold
    repetition_penalty=1.05,  # discourage repeated audio patterns
    max_new_tokens=512,       # max generation steps (1 step = 1/12 sec audio)
)

HTTP API (cross-language integration)

Run BabelVox as an HTTP server so any language (JavaScript, Go, Rust, etc.) can call it:

babelvox --serve --int8 --cp-kv-cache --port 8765

Then from any client:

curl -X POST http://localhost:8765/tts \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello world", "language": "English"}' \
  -o hello.wav

From JavaScript:

const res = await fetch("http://localhost:8765/tts", {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({ text: "Hello world" }),
});
const blob = await res.blob();
const audio = new Audio(URL.createObjectURL(blob));
audio.play();

Endpoints:

Method	Path	Description
POST	/tts	Synthesize speech — JSON body in, WAV bytes out
GET	/health	Health check — returns {"status": "ok"}

POST /tts request body:

Field	Required	Default	Description
text	yes	—	Text to synthesize
language	no	"English"	One of the 10 supported languages
ref_audio	no	null	Path to reference WAV for voice cloning
ref_text	no	null	Transcription of reference audio (improves cloning)
max_new_tokens	no	512	Max generation steps
temperature	no	0.9	Sampling temperature
top_k	no	50	Top-k sampling
top_p	no	1.0	Nucleus sampling threshold
repetition_penalty	no	1.05	Penalty for repeated tokens

Pre-download models

Cache models ahead of time instead of on first use:

from babelvox import download_models
path = download_models()  # downloads ~2.5 GB to HuggingFace cache

API reference

BabelVox(model_path, export_dir, device, precision, ...)

Parameter	Default	Description
model_path	"Qwen/Qwen3-TTS-12Hz-0.6B-Base"	HuggingFace model (tokenizer only)
export_dir	None (auto-download)	Path to exported OpenVINO models
device	"CPU"	"CPU" or "NPU"
precision	"fp16"	"fp16", "int8", "int4", or "fp32"
use_cp_kv_cache	False	KV cache for code predictor (recommended)
talker_buckets	None	NPU bucket sizes, e.g. [64, 128, 256]
cache_dir	None	OpenVINO compiled model cache directory

tts.generate(text, language, ref_audio, ...) returns (waveform, sample_rate)

Parameter	Default	Description
text	required	Text to synthesize
language	"English"	One of the 10 supported languages
ref_audio	None	Path to reference WAV for voice cloning
ref_text	None	Transcription of the reference audio (improves cloning)
speaker_embed	None	Pre-extracted speaker embedding (numpy array)
max_new_tokens	512	Max generation steps (12 steps = 1 sec audio)
temperature	0.9	Sampling temperature (0 = greedy)
top_k	50	Top-k sampling
top_p	1.0	Nucleus sampling threshold
repetition_penalty	1.05	Penalty for repeated tokens

tts.generate_stream(text, language, ..., chunk_frames=12) — same args as generate(), plus:

Parameter	Default	Description
chunk_frames	12	Codec frames per chunk when not using silence detection (12 = ~1 sec)
split_on_silence	False	Yield at natural pauses between words instead of fixed intervals
min_chunk_frames	6	Minimum frames before considering a silence split (~0.5s)
max_chunk_frames	48	Force yield after this many frames even without silence (~4s)
silence_threshold	0.02	RMS energy threshold for silence detection
crossfade_samples	1200	Overlap samples at chunk edges for click-free joins (50ms at 24kHz)

Yields (waveform_chunk, 24000) tuples as audio is generated.

tts.extract_speaker_embedding(audio_path) returns numpy array (1, 1024)

tts.default_speaker — set to a speaker embedding for consistent voice across calls

Exporting models yourself (optional)

The pre-built INT8 models are downloaded automatically. If you want to export from scratch (e.g., for a different quantization), the export scripts in tools/ require PyTorch:

pip install torch qwen-tts nncf
python tools/export_tts_lm.py
python tools/export_speaker_encoder.py
python tools/export_decoder.py
python tools/export_tokenizer_encoder.py
python tools/export_cp_kvcache.py
python tools/export_weights.py
python tools/quantize_models.py --int8

Performance

Optimization progression

Optimization	RTF	Per-step	Notes
FP16 NPU baseline	3.0x	246 ms	Full-recompute, padded to 256 tokens
+ INT8 quantization	2.1x	156 ms	NNCF INT8_SYM weight compression
+ CP KV cache	1.4x	106 ms	Eliminates redundant code predictor recomputation
+ Multi-bucket talker	1.0x	~80 ms	Dynamically picks smallest NPU shape per step

RTF = Real-Time Factor. RTF=1.0x means generating 1 second of audio takes 1 second of compute.

Where the time goes (INT8 + CP KV cache, 256-token bucket)

Component	Device	Time	Share
Talker (28-layer transformer)	NPU	61 ms	57%
Code predictor (15 groups)	CPU	45 ms	43%
Numpy overhead (embeddings, sampling)	CPU	<1 ms	<1%

Multi-bucket scaling

The talker scales linearly with sequence length on NPU. Pre-compiling at multiple sizes and routing to the smallest bucket that fits dramatically reduces wasted compute:

Bucket size	Talker time	Total (+ 45ms CP)	Effective RTF
64	15 ms	60 ms	0.72x
128	22 ms	67 ms	0.80x
192	31 ms	76 ms	0.91x
256	43 ms	88 ms	1.06x

Hardware tested

• CPU: Intel Core Ultra 7 258V (Lunar Lake)
• NPU: Intel AI Boost (~13 TOPS)
• RAM: 32 GB LPDDR5x
• Device: Samsung Galaxy Book5 Pro

Architecture

Qwen3-TTS uses 5 model components orchestrated in an autoregressive loop:

Text --> Tokenizer --> Text Embeddings --> Talker (28L transformer) --> Codec code_0
                                               |
                       Speaker Embedding ------+    code_0 --> Code Predictor (5L) --> codes 1-15
                       (from reference audio)            \--> repeat 15x with KV cache
                                                                     |
                                           All 16 codes --> Tokenizer Decoder --> Waveform

Component	Layers	Hidden	Heads	Device	INT8 size
Talker	28	1024	16Q/8KV	NPU	444 MB
Code predictor	5	1024	16Q/8KV	CPU	79 MB
Tokenizer encoder	--	--	--	NPU	48 MB
Tokenizer decoder	--	--	--	NPU	114 MB
Speaker encoder	--	--	--	NPU	9 MB

CLI reference

Flag	Default	Description
--device	CPU	CPU or NPU
--int8	off	Use INT8 quantized models
--precision	fp16	fp16, int8, int4, or fp32
--cp-kv-cache	off	KV cache for code predictor (recommended)
--talker-buckets	none	Comma-separated NPU bucket sizes (e.g. 64,128,256)
--kv-cache	off	KV cache for talker (not recommended on NPU)
--cache-dir	none	OpenVINO compiled model cache directory
--max-tokens	200	Maximum generation steps
--max-talker-seq	256	Fixed talker padding (when not using buckets)
--max-decoder-frames	256	Max codec frames for audio decoder
--max-kv-len	256	KV cache buffer size (if --kv-cache)
--text	demo text	Text to synthesize
--language	English	Language for synthesis
--ref-audio	none	Reference audio for voice cloning
--ref-text	none	Transcription of reference audio (improves cloning)
--serve	off	Start HTTP server instead of generating once
--host	0.0.0.0	Server bind address
--port	8765	Server port
--output / -o	output.wav	Output WAV file path
--export-dir	auto-download	Directory with exported models (downloads from HuggingFace if not set)
--model-path	Qwen/Qwen3-TTS-12Hz-0.6B-Base	HuggingFace model (tokenizer)

Acknowledgments

Based on Qwen3-TTS by Alibaba Qwen Team (Apache-2.0).

License

Apache-2.0