Tacotron 2 scored a Mean Opinion Score (MOS) of 4.53 out of 5 — within 0.05 points of professionally recorded human speech at 4.58, based on Google’s original 2018 evaluation. Eight years after its release, the model still anchors production TTS pipelines and research benchmarks. The text-to-speech market hit an estimated $4.8 billion to $5.7 billion in 2026, and Tacotron 2’s architecture remains a reference point across that industry. This page covers the latest Tacotron 2 statistics for 2026, including performance benchmarks, developer adoption, voice synthesis market data, and user feedback on neural TTS quality.
Tacotron 2 Statistics 2026 — TL;DR
Tacotron 2 achieved a MOS of 4.53, compared to 4.58 for natural human speech, based on the original Google evaluation published in 2018.
NVIDIA’s PyTorch implementation of Tacotron 2 has accumulated over 5,300 stars and 1,400 forks on GitHub as of mid-2026.
The global text-to-speech market is valued between $4.8 billion and $5.7 billion in 2026, depending on the research firm, with projected growth to $7.9–$35.3 billion by the early 2030s.
In blind tests using 2026-generation voice synthesis, 71% of listeners could not reliably tell AI-generated speech from a human voice, according to a University of Michigan HCI Lab study.
FastSpeech 2 matches Tacotron 2 on MOS while generating mel-spectrograms at roughly 0.02 real-time factor — about 50x faster at inference.
What Is Tacotron 2 and How Does It Work?
Tacotron 2 is a neural network architecture developed at Google for speech synthesis directly from text. It was introduced in a December 2017 paper titled “Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions.” The system has two stages: a recurrent sequence-to-sequence model that maps character embeddings to mel-scale spectrograms, and a modified WaveNet vocoder that converts those spectrograms into audio waveforms at 24 kHz.
The model removed the need for hand-engineered linguistic features like phoneme alignments and prosody models. It trains end-to-end on paired text-audio data. NVIDIA released an open-source PyTorch implementation with distributed training and mixed-precision support, trained on the LJSpeech dataset — a public corpus of roughly 24 hours of English speech.
Tacotron 2 MOS Score vs. Human Speech
The original paper reported a MOS of 4.53 for Tacotron 2, against 4.58 for professionally recorded speech. That 0.05-point gap was the closest any TTS system had come to human-level quality at the time. A separate project on Papers With Code confirmed the same figures.
| System | MOS Score (5-Point Scale) |
|---|---|
| Human Speech (Professional Recording) | 4.58 |
| Tacotron 2 + WaveNet | 4.53 |
| Tacotron 1 | 3.82 |
| Parametric TTS (Baseline) | ~3.50 |
Source: Google, ICASSP 2018 / Papers With Code
How Does Tacotron 2 Compare to Newer TTS Models?
FastSpeech 2, a non-autoregressive model from Microsoft Research, matches Tacotron 2 on MOS in controlled evaluations while running inference roughly 50x faster. FastSpeech 2 achieves a real-time factor of approximately 0.02 on V100 GPUs, generating one second of audio in about 20 ms. Tacotron 2’s autoregressive decoder makes it slower at inference, though its MOS remains competitive.
VITS, an end-to-end model that generates waveforms directly from text without a separate vocoder, produces speech that multiple comparative studies rate as more natural than Tacotron 2, with lower tone error rates. A Springer-published study comparing VITS, Tacotron 2, and FastSpeech 2 found all three among the most widely used TTS systems, but noted VITS and FastSpeech 2 handle non-autoregressive generation more efficiently.
| Model | Architecture | Vocoder Required | Inference Speed | Approx. MOS |
|---|---|---|---|---|
| Tacotron 2 | Autoregressive Seq2Seq | Yes (WaveNet/WaveGlow) | Slower than real-time | 4.53 |
| FastSpeech 2 | Non-Autoregressive | Yes (Parallel WaveGAN) | ~50x real-time | ~4.50 |
| VITS | End-to-End VAE + GAN | No | Real-time | ~4.55 |
| Parallel Tacotron 2 | Non-Autoregressive | Yes | Faster than Tacotron 2 | Rated higher than baseline |
Source: Microsoft Research / Springer / INTERSPEECH 2021
Tacotron 2 Developer Adoption and GitHub Statistics
NVIDIA’s official Tacotron 2 repository on GitHub has over 5,300 stars and 1,400 forks. PyTorch Hub lists the model at 14,800 interactions (a combined stars metric across the broader Deep Learning Examples collection). The model checkpoint hosted on NVIDIA’s NGC Catalog is 107.44 MB, trained on LJSpeech for 1,200 epochs across 8 V100 GPUs with a batch size of 384.
On GitHub Topics, “tacotron2” returns repositories in Python, C++, and Jupyter Notebook — with derivative projects spanning Japanese, Korean, Vietnamese, Brazilian Portuguese, Mizo, and Kreole TTS. Several of these received updates as recently as January 2026, showing continued community activity even as newer architectures gain ground.
| Metric | Value |
|---|---|
| NVIDIA/tacotron2 GitHub Stars | 5,300+ |
| NVIDIA/tacotron2 GitHub Forks | 1,400+ |
| PyTorch Hub Interactions | 14,800+ |
| NGC Model Checkpoint Size | 107.44 MB |
| Training Dataset | LJSpeech (~24 hrs) |
| Training Duration | 1,200 epochs, 8x V100 GPUs |
| Languages with Community Implementations | 10+ |
Source: GitHub / PyTorch Hub / NVIDIA NGC Catalog
Voice Synthesis Market Size in 2026
Published estimates for the global text-to-speech market in 2026 range from $4.1 billion (Business Research Insights) to $5.83 billion (The Business Research Company). The variation reflects different segment definitions — some reports isolate TTS software, others include services, APIs, and developer tools. Across all estimates, the market grew between 10% and 22% year-over-year from 2025.
The broader AI voice generator market was valued at $4.16 billion in 2025 by MarketsandMarkets and is projected to reach $20.71 billion by 2031 at a 30.7% CAGR. Voice cloning alone accounts for an estimated $2.4 billion in 2025. North America holds the largest regional share at approximately 40%, according to multiple analyst reports.
| Source | 2025 Estimate | 2026 Estimate | Projected (2030–2035) | CAGR |
|---|---|---|---|---|
| Global Market Insights | $4.80B | $5.70B | $35.3B (2035) | 22.4% |
| The Business Research Company | $4.92B | $5.83B | $11.49B (2030) | 18.5% |
| MarketsandMarkets (AI Voice Gen) | $4.16B | ~$5.4B | $20.71B (2031) | 30.7% |
| Business Research Insights | $3.65B | $4.10B | $11.1B (2034) | 12.3% |
Source: Global Market Insights / TBRC / MarketsandMarkets / BRI, 2026
Tacotron 2 Statistics on User Feedback and Voice Naturalness
A University of Michigan HCI Lab study found that 71% of callers could not reliably distinguish between AI and a human receptionist in blind tests using 2026-generation voice synthesis. That figure is directional for the broader neural TTS field, not Tacotron 2 specifically, but it reflects the quality trajectory that Tacotron 2 initiated in 2017–2018.
A ScienceDirect study published in October 2025 tested human-AI voice discrimination: participants correctly identified human voices 86% of the time but managed only 55% accuracy on AI-generated voices — meaning synthetic voices “fooled” listeners nearly half the time. Separately, 54% of TTS end-users cited a lack of natural-sounding voices as a barrier to adoption, per Business Research Insights. AWS led the TTS market with 4.2% share in 2025, followed by Google and Microsoft.
| Metric | Value | Source |
|---|---|---|
| Listeners unable to distinguish AI from human voice | 71% | Univ. of Michigan HCI Lab, 2025 |
| Correct identification rate for human voices | 86% | ScienceDirect, Oct 2025 |
| Correct identification rate for AI voices | 55% | ScienceDirect, Oct 2025 |
| End-users citing lack of natural voices as a barrier | 54% | Business Research Insights, 2024 |
| Enterprises using TTS for accessibility | 68% | Business Research Insights, 2024 |
Source: University of Michigan / ScienceDirect / Business Research Insights
Voice AI Market Adoption in 2026
The global voice AI market crossed $22 billion in 2026, according to Ringly.io, which compiled data from Gartner, Statista, and Market.us. Gartner projects $80 billion in contact center labor cost savings from conversational AI this year. An estimated 34% of U.S. small and mid-size businesses have deployed or are piloting AI voice technology as of early 2026.
Consumer adoption data: 157.1 million people in the United States are expected to use voice assistants in 2026, per Statista. There are 8.4 billion voice-enabled devices in use worldwide. Google Assistant leads with about 92.4 million U.S. users, followed by Siri at 87 million and Alexa at 77.6 million, according to eMarketer. Half of all consumers have made a purchase using a voice assistant.
| Metric | Value |
|---|---|
| Global Voice AI Market (2026) | $22.5 billion |
| Contact Center AI Cost Savings (2026) | $80 billion (Gartner) |
| U.S. Voice Assistant Users (2026) | 157.1 million |
| Voice-Enabled Devices Worldwide | 8.4 billion |
| U.S. SMBs Deploying/Piloting Voice AI | 34% |
| Consumers Who Purchased via Voice Assistant | 50% |
Source: Ringly.io / Gartner / Statista / eMarketer / Market.us, 2026
Tacotron 2 Statistics: Known Limitations
Google’s original blog post acknowledged several unresolved issues. Tacotron 2 struggles with complex or uncommon words — “decorum” and “merlot” were cited as examples. In rare cases, the model generates random noise artifacts. It cannot produce audio in real time due to its autoregressive decoder. Emotion or style control is not built into the base architecture.
The over-smoothness problem was documented in a 2019 MDPI paper on Es-Tacotron2. Because Tacotron 2 is a statistical model, it tends to produce “averaged” speech that sounds less expressive than natural recording. Parallel Tacotron 2, published at INTERSPEECH 2021, addressed both speed and robustness by replacing the autoregressive decoder with a non-autoregressive design and adding differentiable duration modeling. That variant eliminates word-skipping and repetition errors common with soft attention.
Tacotron 2 Statistics in Research and Text-to-Speech Applications
The original Tacotron 2 paper has been cited thousands of times across Semantic Scholar, Google Scholar, and IEEE Xplore. It is referenced in nearly every major TTS paper published since 2018, including FastSpeech, FastSpeech 2, VITS, Glow-TTS, and Parallel Tacotron 2. Research groups in more than a dozen countries have adapted Tacotron 2 for low-resource languages, including Vietnamese, Mizo, Moroccan Berber, Czech, Korean, and Japanese.
NVIDIA distributes pre-trained Tacotron 2 checkpoints through its NGC Catalog and PyTorch Hub. The model is part of NVIDIA’s NeMo toolkit and was previously integrated into the Riva (formerly Jarvis) speech AI platform. Training a new voice from scratch requires an NVIDIA GPU with CUDA and cuDNN, the LJSpeech dataset or equivalent, and the open-source codebase. Per NVIDIA’s documentation, speech becomes intelligible around 4,750 training steps and produces high-quality output by 15,750 steps.
How Many Developers Use Tacotron 2 in 2026?
Exact developer counts are not published, but proxy metrics give a picture. NVIDIA’s repository alone accounts for 5,300+ stars. The Rayhane-mamah Tensorflow implementation is another popular fork. Across all “tacotron2” tagged repositories on GitHub, there are dozens of active projects in Python, C++, and Jupyter Notebook. Community-maintained projects received updates through at least January 2026.
Production use has shifted toward faster models. Teams needing real-time inference increasingly choose VITS or FastSpeech 2 over Tacotron 2. Tacotron 2 remains the go-to for offline audio generation — audiobooks, dubbing, and content creation platforms — where latency tolerance is higher and MOS quality is the priority.
Tacotron 2 Statistics: Key Technical Specs
| Parameter | Detail |
|---|---|
| Character Embedding Dimension | 512 |
| Encoder Convolution Layers | 3 |
| Encoder Bi-LSTM Units | 512 |
| Decoder Type | Autoregressive LSTM |
| Audio Output Sample Rate | 24 kHz (WaveNet) / 22,050 Hz (WaveGlow) |
| NGC Checkpoint Size | 107.44 MB |
| Regularization | Dropout (instead of Zoneout) |
| Mixed Precision Training | Supported (FP16 via NVIDIA Apex) |
Source: NVIDIA / Google ICASSP 2018 Paper
What’s Next for Voice Synthesis Beyond Tacotron 2?
The field has moved toward diffusion-based and codec-based architectures. Models like Bark, Tortoise-TTS, and various VALL-E variants generate speech from discrete audio tokens rather than mel-spectrograms, opening up zero-shot voice cloning from seconds of reference audio. The AI voice generator market at $4–5 billion in 2025 is projected at $20+ billion by 2031.
Meta’s acquisition of PlayHT (PlayAI) in July 2025 signals increasing consolidation. AWS, Google, and Microsoft compete on language coverage and voice realism in their cloud TTS APIs. Tacotron 2’s two-stage pipeline — text to spectrogram, spectrogram to waveform — has largely given way to single-stage end-to-end models. But Tacotron 2’s MOS benchmark and architectural ideas still inform new model development.
FAQ
What MOS score did Tacotron 2 achieve?
Tacotron 2 achieved a Mean Opinion Score of 4.53 on a 5-point scale, compared to 4.58 for professionally recorded human speech, based on Google’s 2018 evaluation.
Is Tacotron 2 still used in 2026?
Yes. NVIDIA’s repository has 5,300+ GitHub stars and community projects received updates through January 2026. It remains common in offline audio production where latency is less of a concern.
How big is the text-to-speech market in 2026?
Published estimates for 2026 range from $4.1 billion to $5.83 billion depending on the source and segment definition. Growth rates sit between 10% and 22% year-over-year.
How does Tacotron 2 compare to VITS and FastSpeech 2?
All three produce comparable MOS scores. VITS and FastSpeech 2 run much faster at inference — FastSpeech 2 at roughly 50x real-time — while Tacotron 2’s autoregressive decoder is slower.
Can listeners tell AI voices from human voices?
A 2025 University of Michigan study found 71% of callers could not reliably distinguish AI from human speech. A separate ScienceDirect study reported only 55% accuracy in identifying AI voices.
Sources:
https://arxiv.org/abs/1712.05884
https://ainora.lt/blog/voice-ai-statistics-market-data-2026
https://flauntaudio.com/how-big-is-ai-voice-market-2026/
https://www.gminsights.com/industry-analysis/text-to-speech-market
