AI Voices Sound Entirely Human. What's Next?

🎯 The Question

In 2021, human listeners in the Blizzard Challenge could no longer tell AI speech from a real voice actor. Naturalness is solved — so "how do you make and evaluate expressive voice models?" That's the question the whole field is now grappling with. 0:00

Lord Kelvin (19th century): "When you cannot express it in numbers, your knowledge is of a meager and unsatisfactory kind." — if you can't measure it, you can't improve it.

🏆 The Blizzard Challenge

The most rigorous historical TTS benchmark. Each year they release a few hours of a paid actor's recordings and invite researchers to synthesize that voice; submissions scored by human raters on the Mean Opinion Score (MOS) — 1 (completely unnatural) to 5 (completely natural). 2:32

The calibration trick: organizers snuck real actor recordings in among the generated samples to establish a ceiling. In 2008 the best submission scored 85% of the actor ceiling. By 2021, it scored 99% — the benchmark saturated. "Not that speech generation was solved — the evaluation methods became too rudimentary."

The core problem: "naturalness" is one-dimensional. "It doesn't sound natural" is like a restaurant review — okay, but what does that even mean?

🎭 Naturalness vs Expressiveness

The new buzzword is expressiveness — and every lab defines it differently. 5:13

Inworld's definition: variation and dynamic range across prosodic features. Prosody = intonation, stress, rhythm, loudness.

Hume's decomposition (voice EQ benchmark): affect, pragmatics, prosodic emphasis, sarcasm, emotional transitions, and "production-ready expressive tone."

👥 Evaluation via Humans (MOS)

The core metric is still MOS — a 1–5 average. Simplicity isn't the problem; dependence on logistics is (how raters are selected, environment, instructions). 6:40

The ITU standard: MOS was standardized for telephone-call quality in the '90s by the International Telecommunication Union — "the talker should be seated in a quiet room with volume between 30 and 120 m³ and reverberation <500ms." It's been updated twice for crowdsourcing (Mechanical Turk), but it's unclear how strictly labs follow it today — so MOS scores across labs aren't directly comparable.

Modern practice: collect multiple MOS scores for different aspects (naturalness, expressiveness, role-fit, target-voice similarity, acoustic quality) — a nuanced methodology that doesn't nominate a single winner.

⚔️ Speech Arenas & Elo

The arena format (originally popularized by LLMs) arrived in speech — most notably Artificial Analysis. You hear two snippets and pick a preference; results are ranked by Elo (the chess rating system). 8:38

Relative vs absolute judgment: arenas force relative judgments (better for people who don't know what "1 star" means in isolation), but they average someone else's subjective opinion — and have use-case coverage gaps. "A customer-support voice that's too happy is totally fine on the arena, where people prefer warm voices — but in real setup you'd want someone to speak naturally without extra emotions."

🤖 AI as a Judge

Can AI predict human ratings? The VoiceMOS Challenge (2022) built models that predict MOS from audio. The winner, UTMOS (U. Tokyo), became widely used — even as a data-curation filter (NVIDIA MIMO curator: filter >3.5 for general TTS, >4 for voice cloning) and as a reward signal. DNSMOS (Microsoft's Deep Noise Suppression challenge) does the same for noise. 10:32

But UTMOS/DNSMOS are narrow — naturalness and noise only, not emotion/sarcasm/similarity. The newer speech/audio language models (Gemini, GPT Audio, Kimi Audio) could act as general judges. Hume tested ~20 TTS models with human vs AI judges: correlation was high on language stability (a correctness measure) but dropped below 0.2 on "acting" (highly subjective).

Igor Poletaev (Inworld CSO): "How much do you trust speech language models to give you scores?" → "We don't. Human ears are the best instruments you can have. Speech LMs make a lot of mistakes — they hear things that don't exist, and they don't hear popping sounds that are clear to human ears."

📐 Deterministic Proxy (DSP)

For quick internal needs, Inworld reverse-engineered a deterministic proxy from human labels: average three acoustic features — pitch range, pitch variation, loudness variation. 15:03

Pure digital signal processing (DSP) — no neural networks, runs on any local device, easy to debug and visualize. A plot of expressiveness across two snippets correctly predicts which sounds more expressive.

🏋️ Training for Expressiveness

Three stages: pre-training (1M+ hours = ~115 years of audio, expressiveness implicitly absorbed by next-token prediction), supervised fine-tuning (explicit — markup tags for coughing/laughing/emotions), and alignment to human preferences. 16:23

The Inworld TTS-2 insight — subtitles: emotion classifiers (happy/sad buckets) struggled on real, blended emotions. But "LLMs already annotate context in square brackets — it's essentially subtitles. If LLMs are so good at subtitles, why can't we adopt that?" → Inworld trained TTS-2 to follow subtitle-like inline indications for what a voice should sound like (e.g. a Terminator "hasta la vista, baby" delivery).

DPO for speech: the standard alignment algorithm (from A/B human judgments). The first speech application was Emo-DPO (2024) — instilling emotion via preference pairs from the Emotional Speech Database; human raters preferred the aligned model 89% of the time.

⚠️ Goodhart's Law & Reward Hacking

"When a measure becomes a target, it ceases to be a good measure." Automated-metric alignment can backfire. 21:29

Experiment	What went wrong
Optimizing for intelligibility (CER)	Pitch variability collapsed — flat, monotonous speech (vanilla monotonic speech transcribes more reliably, so expressiveness got implicitly penalized)
Adding speaker-similarity reward	Reward hacking: the model rambled forever, refusing to emit the end-of-sequence token (longer outputs coincidentally scored higher on speaker similarity)
Human preferences (DPO, 200 examples)	Maintained pitch variability and won human blind testing against both the unaligned baseline and the GRPO experiments

"RL with automated metrics only improves the specific metrics you optimize for — hoping for orthogonal improvements as a byproduct is wishful thinking."

🧠 The Hawking Lesson

Expressiveness isn't universally desirable. After ALS robbed Stephen Hawking of speech in 1985, he used a flat, mechanical "Perfect Paul" voice — and kept it for 30 years, refusing newer natural voices. 24:29

"I keep it because I have not heard a voice I like better, and because I have identified with it." — He didn't need to sound human to move humanity.

🔗 The Blizzard Challenge · VoiceMOS Challenge · Artificial Analysis Speech Arena · Voice Lens (DSP proxy) · UTMOS · Emo-DPO

💡 Key Takeaways

Naturalness is saturated — the Blizzard Challenge hit 99% of the actor ceiling in 2021. The new frontier is expressiveness.
Expressiveness is hard to define — Inworld: prosodic variation (intonation/stress/rhythm/loudness); Hume: affect, pragmatics, sarcasm, emotional transitions.
MOS is here to stay but not comparable across labs — it depends on logistics and the (loosely followed) ITU standard.
Arenas give relative judgments (Elo) but have use-case coverage gaps (a "too happy" support voice ranks well but is wrong in production).
AI judges (UTMOS/DNSMOS) are narrow — and speech LMs "make mistakes; human ears are the best instrument."
A deterministic DSP proxy (pitch range + pitch variation + loudness variation) gives a fast, debug-able expressiveness signal.
Inworld's breakthrough: subtitle-style direction — train the model to follow inline indications instead of emotion buckets.
Goodhart's law is real — automated-metric RL flattened prosody and caused reward hacking; human-preference DPO won blind tests.
Expressiveness isn't always the goal — Hawking kept his flat voice for 30 years.

📍 Timestamp Index

0:00 Intro

2:32 The Blizzard Challenge

5:13 Naturalness vs expressiveness

6:40 Evaluation via humans (MOS)

8:38 Speech arenas & Elo

10:32 AI as a judge

15:03 Deterministic proxy (DSP)

16:23 Training for expressiveness

21:29 Goodhart's law & reward hacking

24:29 The Hawking lesson

AI Voices Sound Entirely Human. What's Next? — Inworld TTS-2