Quality Evals
Dubverse vs Competitors: Text-to-Speech Comparison
This document provides a detailed comparison of Dubverse with other Text-to-Speech (TTS) providers, including audio samples and specific observations.
Hindi Audio Comparison
Key Observations
Dubverse
- Clear pronunciation with minor issues (e.g., “टेस्टी” slightly unclear)
- Consistent speed and natural-sounding speech
- Good audio quality
- Handles complex Hindi sentences well
Competitors
- ElevenLabs: Mispronunciations, slow speed
- XTTS: Pronunciation issues, stuttering, inconsistent audio quality
- Sarvam: Glitchy audio, missed words, no English support
- Bhashini AI4Bharat: Poor audio quality, fast speed, unclear pronunciations
- Bhashini IITM: Fast audio, pronunciation issues
- Cartesia: Missing words, fast speed, robotic sound
- PlayHT: Slow speed, lacks emotion
- MicMonster: Electronic sound, unnatural speech
English Audio Comparison
Key Observations for English
Dubverse
- Natural-sounding speech
- Appropriate speed and intonation
- Handles questions and statements well
Competitors
- ElevenLabs: Hallucination for short sentences
- XTTS: Noisy audio with poor quality
- Sarvam: No English support
- Bhashini AI4Bharat: No English support
- Bhashini IITM: No English support
- Cartesia: Robotic sound, mispronunciations
- PlayHT: Too slow, lacks emotion
- MicMonster: Electronic sound, unnatural
Emotional Sentences Test
| Sentence | Audio |
|---|---|
| I can’t believe it! This is amazing! | |
| Oh my gosh, did you see that? |
Note: ElevenLabs demonstrates a high pitch issue, especially for female voices, which can sound unnatural in emotional contexts.
Why Choose Dubverse?
-
Superior Hindi Support: Dubverse outperforms competitors in handling complex Hindi sentences with clear pronunciation and natural intonation.
-
Multilingual Capabilities: Unlike some competitors, Dubverse excels in both Hindi and English, making it ideal for multilingual projects.
-
Consistent Quality: Dubverse maintains high audio quality across different sentence types and languages.
-
Natural Speech Patterns: Our TTS technology closely mimics human speech patterns, avoiding the robotic or electronic sound common in other solutions.
-
Emotional Range: While competitors struggle with emotional sentences, Dubverse can convey a wide range of emotions naturally.
-
Balanced Speed: Dubverse strikes the right balance between clarity and natural speech speed, unlike competitors that are either too slow or too fast.
-
Versatility: From simple greetings to complex tongue-twisters, Dubverse consistently delivers high-quality speech synthesis.
Open Source Models
Zonos
Zonos is an open-weight text-to-speech model trained on over 200k hours of multilingual speech data. While it shows promising capabilities, our evaluation reveals some important limitations to consider before production use.
Key Features
- Zero-shot voice cloning with 10-30s speaker samples
- Multilingual support (English, Japanese, Chinese, French, German)
- Fine-grained control over speaking rate, pitch, audio quality and emotions
- Real-time factor of ~2x on RTX 4090
Evaluation Results
Our testing revealed several areas that need improvement before production deployment:
- Audio Quality
- Inconsistent speaker stability with occasional hallucinations
- Voice characteristics sometimes drift during longer utterances
- Generation errors can cause interruptions in output
- Expressiveness
- While capable of emotional speech, results are inconsistent
- American male samples showed unintended “happy, chirpy tone”
- Faster speaking pace than intended in some dialogues
- Hallucination Issues
- More prevalent in American male speaker compared to other voices
- Can manifest as unexpected voice changes mid-speech
- Affects overall reliability of the output
While Zonos demonstrates impressive capabilities for an open source model, these limitations make it better suited for experimental use rather than production applications where consistent, reliable output is critical.
Audio Samples
Good Narration Examples
| Description | Audio |
|---|---|
| Female Voice Narration | |
| Calm Voice Narration |
Quality Issues
| Description | Audio |
|---|---|
| Wrong Tonality | |
| Robotic Sound |
Hallucination Examples
| Description | Audio |
|---|---|
| Hallucination 1 | |
| Hallucination 2 | |
| Hallucination 3 | |
| Hallucination 3 (with male voice) | |
| Hallucination 4 |
Voice Samples
| Description | Audio |
|---|---|
| American Male |
Conclusion
Dubverse’s candy.two model stands out as a superior choice for text-to-speech needs, especially for projects requiring high-quality Hindi and English voice synthesis. With its natural-sounding speech, consistent performance across languages, and ability to handle complex sentences, candy.two offers a robust solution that outperforms many established competitors in the market. For more details on candy.two and our other TTS models, check out our Models Overview.