Datasets:
audio audioduration (s) 1.04 19.1 | text stringlengths 5 329 | ref_audio stringclasses 1
value |
|---|---|---|
In this report, we present the QEN3-TTS series, | data/ref_audio.wav | |
In this report, we present the QEN3-TTS series, | data/ref_audio.wav | |
A family of advanced, multilingual, controllable, robust, and streaming text-to-speech models. | data/ref_audio.wav | |
A family of advanced, multilingual, controllable, robust, and streaming text-to-speech models. | data/ref_audio.wav | |
Quen 3 TTS supports state-of-the-art 3-second voice cloning and description-based control, allowing both the creation of entirely novel voices and fine-grained manipulation | data/ref_audio.wav | |
Quen 3 TTS supports state-of-the-art 3-second voice cloning and description-based control, allowing both the creation of entirely novel voices and fine-grained manipulation | data/ref_audio.wav | |
over the output speech. Trained on over 5 million hours of speech data spanning 10 languages, | data/ref_audio.wav | |
over the output speech. Trained on over 5 million hours of speech data spanning 10 languages, | data/ref_audio.wav | |
Quan 3 TTS adopts a dual track | data/ref_audio.wav | |
Quan 3 TTS adopts a dual track | data/ref_audio.wav | |
LM architecture for real-time synthesis coupled with two speech tokenizers. One, | data/ref_audio.wav | |
LM architecture for real-time synthesis coupled with two speech tokenizers. One, | data/ref_audio.wav | |
QEM3 TTS Tokenizer is a single codebook codec | data/ref_audio.wav | |
QEM3 TTS Tokenizer is a single codebook codec | data/ref_audio.wav | |
Emphasizing semantic content, which offers seamlessly integration with Quen audio | data/ref_audio.wav | |
Emphasizing semantic content, which offers seamlessly integration with Quen audio | data/ref_audio.wav | |
and enable Streaming Waveform Reconstruction via blockwise to when | data/ref_audio.wav | |
and enable Streaming Waveform Reconstruction via blockwise to when | data/ref_audio.wav | |
TTS tokenizer, 12 Hertz. | data/ref_audio.wav | |
TTS tokenizer, 12 Hertz. | data/ref_audio.wav | |
achieves extreme bitrate reduction and ultra-low latency streaming, enabling immediate | data/ref_audio.wav | |
achieves extreme bitrate reduction and ultra-low latency streaming, enabling immediate | data/ref_audio.wav | |
First packet emission, 97 milliseconds, through its 12.5 hertz, | data/ref_audio.wav | |
First packet emission, 97 milliseconds, through its 12.5 hertz, | data/ref_audio.wav | |
16-layer multi-codebook design and a lightweight casual conv-net. | data/ref_audio.wav | |
16-layer multi-codebook design and a lightweight casual conv-net. | data/ref_audio.wav | |
Extensive experiments indicate state-of-the-art performance across diverse objective and subjective benchmark. | data/ref_audio.wav | |
Extensive experiments indicate state-of-the-art performance across diverse objective and subjective benchmark. | data/ref_audio.wav | |
e.g. TTS multilingual test set instruct TTS eval | data/ref_audio.wav | |
e.g. TTS multilingual test set instruct TTS eval | data/ref_audio.wav | |
And our long speech test set. To facilitate community research and development, we release | data/ref_audio.wav | |
And our long speech test set. To facilitate community research and development, we release | data/ref_audio.wav | |
Both tokenizers and model under the Apache 2.0 license. | data/ref_audio.wav | |
Both tokenizers and model under the Apache 2.0 license. | data/ref_audio.wav | |
One. Introduction. Figure one. | data/ref_audio.wav | |
One. Introduction. Figure one. | data/ref_audio.wav | |
Stable, controllable, and human-like speech synthesis is widely viewed as a key capability on the path to AGI. | data/ref_audio.wav | |
Stable, controllable, and human-like speech synthesis is widely viewed as a key capability on the path to AGI. | data/ref_audio.wav | |
Modern Neural Text-to-Speech models trained on large-scale datasets | data/ref_audio.wav | |
Modern Neural Text-to-Speech models trained on large-scale datasets | data/ref_audio.wav | |
already deliver exceptional capability to generate high-quality speech from a few seconds of reference audio. | data/ref_audio.wav | |
already deliver exceptional capability to generate high-quality speech from a few seconds of reference audio. | data/ref_audio.wav | |
Among them, discrete speech tokenization, combined with autoregressive language modeling of discrete units, has gained traction, offering improved stability. | data/ref_audio.wav | |
Among them, discrete speech tokenization, combined with autoregressive language modeling of discrete units, has gained traction, offering improved stability. | data/ref_audio.wav | |
while preserving high naturalness and human likeness. | data/ref_audio.wav | |
while preserving high naturalness and human likeness. | data/ref_audio.wav | |
Conditioning on vocal features or text instructions facilitates finer grain control over porosity. | data/ref_audio.wav | |
Conditioning on vocal features or text instructions facilitates finer grain control over porosity. | data/ref_audio.wav | |
and style, resulting in outputs of greater richness and diversity. | data/ref_audio.wav | |
and style, resulting in outputs of greater richness and diversity. | data/ref_audio.wav | |
These breakthroughs are paving the way for diverse applications in fields such as virtual assistants | data/ref_audio.wav | |
These breakthroughs are paving the way for diverse applications in fields such as virtual assistants | data/ref_audio.wav | |
and automated content creation. In this report | data/ref_audio.wav | |
and automated content creation. In this report | data/ref_audio.wav | |
We take steps towards stable, controllable, and human-like speech synthesis and introduce QEM3-TTS, | data/ref_audio.wav | |
We take steps towards stable, controllable, and human-like speech synthesis and introduce QEM3-TTS, | data/ref_audio.wav | |
The first text-to-speech model in the Quent series. Quent 3 TTS exhibits the following properties. | data/ref_audio.wav | |
The first text-to-speech model in the Quent series. Quent 3 TTS exhibits the following properties. | data/ref_audio.wav | |
1. Controllability Quen 3 TTS allows users to create new voices or manipulate | data/ref_audio.wav | |
1. Controllability Quen 3 TTS allows users to create new voices or manipulate | data/ref_audio.wav | |
Fine-grained attributes of generated speech | data/ref_audio.wav | |
Fine-grained attributes of generated speech | data/ref_audio.wav | |
via natural language descriptions while also supporting this stable generation | data/ref_audio.wav | |
via natural language descriptions while also supporting this stable generation | data/ref_audio.wav | |
of any content using the created voice. Two, voice cloning and predefined voice profiles | data/ref_audio.wav | |
of any content using the created voice. Two, voice cloning and predefined voice profiles | data/ref_audio.wav | |
Quan 3 TTS supports 3-second voice cloning and generation using a set of X-curated, | data/ref_audio.wav | |
Quan 3 TTS supports 3-second voice cloning and generation using a set of X-curated, | data/ref_audio.wav | |
High Quality Preset Voices 3. Naturalness Beyond achieving our robust synthesis, | data/ref_audio.wav | |
High Quality Preset Voices 3. Naturalness Beyond achieving our robust synthesis, | data/ref_audio.wav | |
Quen 3 TTS excels in generating highly natural and expressive speech. | data/ref_audio.wav | |
Quen 3 TTS excels in generating highly natural and expressive speech. | data/ref_audio.wav | |
Our 1.7b model, in particular, delivers state-of-the-art human-like | data/ref_audio.wav | |
Our 1.7b model, in particular, delivers state-of-the-art human-like | data/ref_audio.wav | |
Quality. Demonstrating our approach successfully maximizes perceptual quality. | data/ref_audio.wav | |
Quality. Demonstrating our approach successfully maximizes perceptual quality. | data/ref_audio.wav | |
without overfitting to ASR-related metrics. 4. Multilinguality | data/ref_audio.wav | |
without overfitting to ASR-related metrics. 4. Multilinguality | data/ref_audio.wav | |
The model is trained across more than 10 languages and supports speaker-consistent multilingual generation. | data/ref_audio.wav | |
The model is trained across more than 10 languages and supports speaker-consistent multilingual generation. | data/ref_audio.wav | |
Streaming. | data/ref_audio.wav | |
Streaming. | data/ref_audio.wav | |
Designed for streaming text input and streaming audio output, it achieves a first packet latency as low as 97 milliseconds. | data/ref_audio.wav | |
Designed for streaming text input and streaming audio output, it achieves a first packet latency as low as 97 milliseconds. | data/ref_audio.wav | |
for the 0.6B variant and 101 milliseconds for the 1.7B. | data/ref_audio.wav | |
for the 0.6B variant and 101 milliseconds for the 1.7B. | data/ref_audio.wav | |
Beyond the aforementioned aspects, | data/ref_audio.wav | |
Beyond the aforementioned aspects, | data/ref_audio.wav | |
And from a broader perspective of practical application, it is crucial | data/ref_audio.wav | |
And from a broader perspective of practical application, it is crucial | data/ref_audio.wav | |
For our model to integrate seamlessly with large language models and achieve extremely low | data/ref_audio.wav | |
For our model to integrate seamlessly with large language models and achieve extremely low | data/ref_audio.wav | |
First Packet Latency To this end, we use discrete speech representations as the cornerstone of our | data/ref_audio.wav | |
First Packet Latency To this end, we use discrete speech representations as the cornerstone of our | data/ref_audio.wav | |
architecture and introduce two tokenizers in the Quen 3 TTS family. One, Quen 3 | data/ref_audio.wav | |
architecture and introduce two tokenizers in the Quen 3 TTS family. One, Quen 3 | data/ref_audio.wav | |
TTS Tokenizer 25Hz | data/ref_audio.wav | |
TTS Tokenizer 25Hz | data/ref_audio.wav | |
employs a 25-hertz single code block representation with waveform reconstruction | data/ref_audio.wav | |
employs a 25-hertz single code block representation with waveform reconstruction | data/ref_audio.wav |
End of preview. Expand in Data Studio
soch-2h TTS Dataset
Generated by dataset-maker.
- Samples: 984
- Format: HuggingFace audiofolder (Qwen 3 TTS compatible)
- Columns:
audio— audio segment (decoded Audio feature, with player in viewer)text— transcriptref_audio— string path to reference speaker audio (data/ref_audio.wav)
- Downloads last month
- 99