What TTS throws away

When a podcast host says "spaaaace" and holds the vowel for 1.1 seconds, every ASR (automatic speech recognition) transcribes it as "space." When two friends banter and one reacts with "Oh. Ooh. Mm." for 2.4 seconds of pure non-verbal communication, the transcript shows three punctuated words. When a speaker abandons a clause mid-thought — "I really have no — I never thought…" — the transcript records an em-dash. The audio records an affect shift.

That's the gap. The training data for every TTS system is built from transcripts that discard the expressive vocabulary humans actually use. Drawn-out vowels, reactive backchannels, mid-sentence register shifts, the timing of a laugh — all of it is flattened into clean text before a TTS model ever sees it. The model can't learn what it's never shown.

This post is about what that gap sounds like, why it exists, and what it would take to close it. Press play below — then expand the comparisons to see what three STT systems and four TTS engines make of the same 25 seconds.

step 2 · what TTS does with that transcriptSame transcript fed to 4 frontier TTS labs — listen, then see where the time goes

The original is 25 seconds. The same 8-segment script rendered through each TTS lab takes 28.7s to 34.9s. The chart aligns all 9 versions on a shared time axis, colored by speaker. Clean = raw STT transcript; Enhanced = same transcript with lab-appropriate affective markup. Switch tabs to compare timing vs silence placement.

Timinghow long each segment tookAmplitude envelopewhere silence vs speech lands

0s

5s

10s

15s

20s

25s

30s

Original audio (Scribe v2)

Oh.

If you watch…

Mm. But for a massive chunk…

That's true.

And then when I went off…

I mean, I really have no —…

[laughs]

But the point…

24.68s

ElevenLabs v3 — clean

Oh.

If you watch…

Mm. But for a massive chunk…

That's true.

And then when I went off…

I mean, I really have no —…

[laughs]

But the point…

28.74s

ElevenLabs v3 — enhanced

Oh.

If you watch…

Mm. But for a massive chunk…

That's true.

And then when I went off…

I mean, I really have no —…

[laughs]

But the point…

29.48s

Gemini 3.1 Flash — clean

Oh.

If you watch…

Mm. But for a massive chunk…

That's true.

And then when I went off…

I mean, I really have no —…

[laughs]

But the point…

31.32s

Gemini 3.1 Flash — enhanced

Oh.

If you watch…

Mm. But for a massive chunk…

That's true.

And then when I went off…

I mean, I really have no —…

[laughs]

But the point…

31.16s

Gemini 2.5 Pro — clean

Oh.

If you watch…

Mm. But for a massive chunk…

That's true.

And then when I went off…

I mean, I really have no —…

[laughs]

But the point…

29.21s

Gemini 2.5 Pro — enhanced

Oh.

If you watch…

Mm. But for a massive chunk…

That's true.

And then when I went off…

I mean, I really have no —…

[laughs]

But the point…

29.41s

OpenAI gpt-4o-mini-tts — clean

Oh.

If you watch…

Mm. But for a massive chunk…

That's true.

And then when I went off…

I mean, I really have no —…

[laughs]

But the point…

34.90s

OpenAI gpt-4o-mini-tts — enhanced

Oh.

If you watch…

Mm. But for a massive chunk…

That's true.

And then when I went off…

I mean, I really have no —…

[laughs]

But the point…

33.96s

speaker_0 (Crissle, F) speaker_1 (Kid Fury, M)hover any bar to see the full segment text · TTS per-segment widths estimated by char-ratio × measured total duration; original uses Scribe v2 word timestamps

Original (24.68s, bolded) is the reference. OpenAI expanded most (34.9s) — stitching seams between speakers. ElevenLabs and Gemini 2.5 Pro closest to the human (~29s). Hover any bar for segment text; press play on any row.

0s

5s

10s

15s

20s

25s

30s

Original audio

24.68s

ElevenLabs v3, clean

28.74s

ElevenLabs v3, enhanced

29.48s

Gemini 3.1 Flash, clean

31.32s

Gemini 3.1 Flash, enhanced

31.16s

Gemini 2.5 Pro, clean

29.21s

Gemini 2.5 Pro, enhanced

29.41s

OpenAI gpt-4o-mini-tts, clean

34.90s

OpenAI gpt-4o-mini-tts, enhanced

33.96s

Amplitude envelopes (ffmpeg showwavespic), same time scale as the Timing tab. The dramatic gaps in OpenAI's strips are stitching seams — silence between concatenated segments rather than natural breath spacing.

That gap is what this project is trying to close. ElevenLabs v3 gives you [cheerful] or [deadpan]. It will not give you the mid-thought interruption above — one speaker abandoning a clause and restarting with softer affect. The transcript reads as a clean sentence with an em-dash. The audio is something no tag captures.

What I'm building

We're building a fine-tuned script writer — a model whose output has the affective contour, the cue tags, the drawn-out vowels, and the per-segment direction baked in, so when any TTS reads it, what comes out is something a human voice actor would recognise as expressive.

The script writer needs to emit four layers of paralinguistic detail that conventional script generation throws away:

• Drawn-out vowels as literal spellings — heyyy, spaaaace, okaaaay. Repeated vowels encode hold duration. Standard ASR normalizes these away; the training corpus has to preserve them.
• Non-verbal vocal cues as inline tokens — uh, mm-hm, hmm, oh, and their held forms (uhhhh, mmmmm). Their placement is the difference between a flat read and a believable one.
• Bracketed event tags — [laughs], [sighs], [breath], [whispers], [shouts] — placed at the exact word boundary where they occur, not as global sentence markers.
• Director's-note metadata at the paragraph level — tone, register, pace, emotion, per-speaker affect — so the same model can produce comedy, meditation, dialogue, or banter with different scene direction.

The writer has to cover the full emotional range — not just happy/sad/angry/neutral but blended states like amused-resigned, hopeful-skeptical, warm-sarcastic. Current TTS can't reach those from a flat transcript alone; a script writer that puts affect where it goes is the leverage point.

"Expressive" is the load-bearing word. The goal isn't "monologue for one tired podcaster." It's the full surface area of human speech — comedy timing, meditation silences, conversational backchannels, scene dialogue, songlike delivery — and under all of those, the emotion spectrum a real voice carries: warmth, amusement, sarcasm, fatigue, grief, hopefulness, curiosity. And the blended states that are most of human affect (amused-resigned, hopeful-skeptical, excited-anxious). The script-writer is the layer that says "here is exactly which emotion goes where in the script", and the downstream TTS gets to focus on rendering instead of guessing.

The script writer starts as a LoRA fine-tune, scaling to a full 7-8B fine-tune if tag placement accuracy demands it. (Fish Audio S2 fine-tuned a full 30B model for their annotator; my task is narrower but LoRA may not have enough capacity for ~110 inline tags.) The downstream TTS can be any engine — the first validation gate is whether tagged scripts produce measurably better output than untagged ones. Portability across TTS engines is the only way to keep iterating without vendor lock-in.

The immediate question: where do the training annotations come from? Hand-labeling 100 hours at this granularity would take months. Frontier audio LLMs (Vertex Gemini 2.5 Pro) do it well but cost money and lock you into a closed API. Self-hosting an audio LLM that gets close enough would let us scale without rate limits, without sending audio to a third party, and with a path to fine-tune the annotator later. That's the build documented here.

TTS in May 2026: what moved, what didn't

As of May 2026, audio LLMs and TTS are in two different places. Audio understanding has moved fast. Frontier closed models — Vertex Gemini 2.5 Pro, OpenAI gpt-realtime — can listen to a 5-minute podcast and produce structured descriptions: speakers, emotions, vocal events, drawn-out words. Open models are catching up at roughly one per quarter — Qwen3-Omni, Kimi-Audio, MiDashengLM. Audio synthesis has also moved. ElevenLabs v3 exposes ~70 inline tags; Hume Octave 2 takes free-text prosodic direction; Google Gemini 3.1 Flash TTS accepts scene-level direction inline. But the ceiling moved, it didn't disappear. Every closed system still applies one uniform style across each tagged region. Continuous within-span affect — sarcasm, the smirk-while-talking — remains unsolved.

The corpus + script-writer project lives in that asymmetry: I can build the training data from today's audio understanding models, and bet that synthesis catches up (or that even current synthesis renders richly-tagged scripts better than flat ones).

The annotation bottleneck nobody talks about

The biggest lever for expressive TTS quality isn't model architecture or training scale — it's annotation quality.

Fish Audio's S2 improved its Tag Activation Rate from 62.6% to 88.1% — a 41% relative improvement — by better-annotating its training data. Same architecture, same pipeline, just richer labels. (arXiv:2603.08823) The DeEAR framework showed a 3x improvement in emotion expressiveness from data curation alone. And SpeechJudge found that a fine-tuned reward model (77.2% accuracy) outperforms every frontier AudioLLM at judging speech quality, while traditional metrics like UTMOS (53.7%) are barely better than coin flips. Annotation quality drives everything — training data and evaluation.

The field is converging on a spectrum of annotation approaches:

• Zero annotation — pure scale: Moshi trained on 7M hours with no labels, relying on a neural codec to preserve signals implicitly. Scale, but no control.
• Emergent features: Bark's [laughs] emerged from co-occurring text + audio events. Works for some events, unreliable for others.
• Automated supervision: NaturalSpeech 3 factorizes audio into content, prosody, timbre, and acoustic detail. Principled, but narrow per axis.
• Rich transcription — inline tags at word positions: Fish S2, NVSpeech, and CosyVoice 3 all converge here. Best control, moderate scale.
• Annotation-as-reward — the annotator also serves as the RL reward signal: Fish S2 and CosyVoice 3 both do this.

We're building in the "rich transcription" tier, aiming toward "annotation-as-reward."

The plan, in one picture

Take 5-minute audio clips, run each through an annotation pipeline that emits structured JSON, use those triples as supervised-learning data. In practice, the dark box is a multi-pass pipeline — VAD, diarization, ASR, emotion classifiers, pitch extractors — because no single audio LLM captures everything. The EACL 2026 finding on "lexical dominance" (Gemini scores 96.6% on text-based audio tasks but 25-35% on audio-only paralinguistic tasks) is why. A noisy annotator means a script writer that puts [laughs] in the wrong places.

The moment that started this benchmark

The first model I reached for was Audio Flamingo Next, NVIDIA's 8B audio model. On the first run, it returned identical output for every clip — same speaker count, same tone, same four breath events at the same timestamps. NPR news, a Moth story, a meditation, a stand-up set, all the same JSON. It wasn't broken. It just wasn't listening; the audio kwarg was being silently dropped by the processor.

Model leaderboards are not model deployments. The number on the HuggingFace page is one promise; the number you get on your data, in your container, is a different one. This post documents the gap.

Calibrating the annotation box is the single highest-leverage decision in the whole pipeline.

Three layers of the gap, with audio

Three clips, three failure modes.

Layer 1 — The "uniform style per region" problem

Same clip. Pay attention to the mid-thought abandon: "And I mean, I really have no — I never thought that there was a chance that you have seen that." That em-dash isn't punctuation — it's an affect shift. Confident anecdote to softer acknowledgment, inside a single breath.

Layer 2 — The "complex blended affect" problem

The opening of a Smartless live show. Three voices pile onto a single beat — "One more crack at it" — then hyped-up gratitude, then a self-deprecating setup about smacking Sean across the face. The same speaker simultaneously performs excitement and undercuts himself with dry self-deprecation. Two affects, same words.

Layer 3 — The "drawn-out vowel" problem

Tara Brach delivering a meditation. Listen to how she pronounces "space" — held for around a second, four times longer than dictionary expects. The silences between phrases run 2-3 seconds. None of this duration information appears in the transcript.

What happens when you feed it back to TTS

Each audio block above has a collapsible TTS comparison: four labs (Gemini 3.1 Flash, Gemini 2.5 Pro, ElevenLabs v3, OpenAI gpt-4o-mini-tts), each in two modes — Clean (raw STT, no tags) and Enhanced (lab-appropriate markup).

Key findings across all four labs:

1. Gemini 2.5 Pro absorbs slow-pace direction most aggressively — Tara Brach expanded from ~16.5s to ~31.7s (nearly 2x) but with 13-24s latency per call vs Flash's ~3-5s.
2. ElevenLabs v3 is second-most-responsive — Tara Brach 14.3s → 25.0s (1.75x expansion); its [laughs] tag landed mid-utterance exactly right on The Read.
3. Gemini 3.1 Flash takes direction less dramatically — Tara Brach 14.2s → 19.6s. Some tags ([softly], [whisper]) tripped the safety filter on the meditation script.
4. Gemini's multi-speaker is hard-capped at 2 speakers via the public API (docs claim more; the API rejects 3+). OpenAI has no native multi-speaker at all.

The delta between clean and enhanced rows is the whole point. Listen to Tara Brach across the four labs — that's where the gap is loudest.

The two-speaker correction. And how the pipeline derived it

On first pass we treated The Read as single-speaker. It's a back-and-forth between Crissle West (F) and Kid Fury (M). The correction shows what the corpus pipeline does at scale: Scribe v2's speaker_id tags break the clip into 8 turns; gender labels are cross-checked (no STT returns gender natively); each turn routes to the appropriate voice per lab. Speaker IDs from one model, gender from another, voice routing from a manual mapping — "STT → faithful TTS" is not closed by any single product.

The accent problem

No production ASR returns accent or dialect information. Deepgram, ElevenLabs, OpenAI, CrisperWhisper — all return a BCP-47 language tag or just text. None returns "Black American", "Southern AAVE", or "British RP". Open-source classifiers exist (CommonAccent covers 16 English accents) but AAVE isn't a class in any of them.

The consequence: The Read's hosts code-switch into AAVE constantly. Feed a clean transcript to a default TTS voice and you get a generic white-American reading. We had to route accent manually — ElevenLabs via voice library, Gemini and OpenAI via preamble direction. Adherence is hit-or-miss. Accent-aware TTS isn't a production capability in May 2026. The end-to-end pipeline "audio → derived accent → matched TTS voice" doesn't exist as a shipped product.

Why no production TTS handles these well

Subtler affects are even further behind. Sarcasm requires modeling the divergence between semantic content and prosodic delivery — no production TTS exposes a controllable knob for it, for irony, or for the polite-chuckle-versus-genuine-laugh distinction. Non-frontier TTS (Bark, Sesame CSM, Kyutai Moshi) sits well behind on emotion expressiveness.

The thread through all three layers: at every current TTS tier, continuous human affect is approximated by discrete styling. Our bet: build a corpus whose training data carries the affective contour finer than the global tag, pairing paragraph-level scene metadata with per-speaker inline markup, so the writer learns when to describe affect versus when to mark it.

The same annotation pipeline that creates the training data can also validate TTS output. Fish Audio S2 and CosyVoice 3 both reuse their annotation models as RL reward signals. Building the annotator isn't just about creating training data — it's about creating the evaluation engine for everything downstream.

The open question: does a script writer trained on richly annotated data produce scripts that downstream TTS renders more expressively than an untagged baseline? That validation is in progress; we'll share tag activation rates and listener preference numbers when they're ready.

Three takeaways:

1. Script quality matters more than model choice. The delta between clean and enhanced TTS output comes from what's in the script, not from switching engines.
2. No single audio LLM captures everything. Drawn-out vowels, mid-sentence affect shifts, and backchannels require specialist acoustic models, not just bigger language models.
3. The annotation pipeline is the product. Fish S2's 41% TAR improvement came from annotation alone. The data is the moat.

If you work in speech and see something I've gotten wrong — a methodology gap, a better tool, a paper I should have cited — I'd genuinely like to know. Reach me at @DavidAmal.