VoxCPM-Kernel/VoxCPM2-Turbo-Listening-Samples

VoxCPM2-Turbo listening samples

Five English sentences spoken by VoxCPM2 through the VoxCPM2-Turbo engine, each in two modes, so you can hear the difference. In each mode the 5 requests were sent at the same moment to one NVIDIA A10.

At a glance

Model	openbmb/VoxCPM2
Engine	VoxCPM2-Turbo
GPU	1× NVIDIA A10
Requests	5 at once, per mode
Modes	default: the original VoxCPM2 algorithm (10 diffusion steps, CFG 2.0) with INT8 math; low_latency: opt-in settings that send the first audio sooner
Voice	none given: the model picks a voice per request, so the two versions of a sentence can sound like different speakers
Audio	10 WAV files, 48 kHz mono, 16-bit
Related	VoxCPM2-Turbo-Benchmark-200-Requests: the same two modes with 200 requests at once, with quality scores

Samples

#	Text	Low-latency	TTFA	Length	Default	TTFA	Length
1	Hello! Welcome to VoxCPM2-Turbo, a fast text-to-speech engine that can serve hundreds of people at the same time.	low_latency/sample_1.wav	1.5 ms	7.04 s	default/sample_1.wav	16.9 ms	7.04 s
2	The weather forecast predicts sunshine for most of the weekend, with a light breeze along the coast.	low_latency/sample_2.wav	1.3 ms	6.4 s	default/sample_2.wav	16.7 ms	3.68 s
3	Please confirm your appointment by replying to this message before Friday afternoon.	low_latency/sample_3.wav	1.2 ms	4.64 s	default/sample_3.wav	16.6 ms	4.16 s
4	Artificial intelligence is changing how we write, code, and communicate with each other every day.	low_latency/sample_4.wav	1.0 ms	5.92 s	default/sample_4.wav	16.5 ms	5.44 s
5	One by one, the campfires were extinguished, and the oasis fell as quiet as the desert.	low_latency/sample_5.wav	0.9 ms	5.28 s	default/sample_5.wav	16.4 ms	4.64 s

• TTFA (time to first audio): from sending the request to receiving its first 160 ms audio chunk. In low_latency that first chunk is mostly silence, so the voice starts a little later.
• All 10 files were transcribed with Whisper large-v3; every transcript matches its sentence.
• Low-latency settings: a faster sampler (CASC ultra), the first chunk computed from the first text token and served from a cache of first chunks, and a scheduler that decodes each stream before its playback would stall.

Contents

Path	What it is
low_latency/sample_<N>.wav	sentence N, low-latency mode
default/sample_<N>.wav	sentence N, default mode
metadata.csv	file_name, mode, text, ttfa_ms, audio_seconds of every file

Load

from huggingface_hub import snapshot_download
folder = snapshot_download("VoxCPM-Kernel/VoxCPM2-Turbo-Listening-Samples", repo_type="dataset")
# folder/low_latency/sample_1.wav, folder/default/sample_1.wav, folder/metadata.csv

License

Apache-2.0, as VoxCPM2, which generated the audio.