Long-form TTS is broken: how TADA and Borderless are fixing it
“TTS demos always use one sentence. Ask yourself why.”
TL;DR
Every TTS system sounds great on a sentence. Stretch to five minutes and you get word skipping, repetition, speaker drift, and hallucinated content. A root cause is the imbalance between text and audio: codec-based TTS takes roughly 4 to 40 times as many generation steps as there are text tokens. TADA addresses this with one-to-one text-acoustic alignment, which rules out skipped or inserted tokens by design and fits about 11 minutes of speech in a 2,048-token window. Borderless Long Speech Synthesis takes the opposite approach: an agent-centric framework that plans prosody, emotion, and speaker interactions using chain-of-thought reasoning before generating audio. One solves it in the model. The other solves it in the scaffolding. For how TTS fits into the real-time voice agent stack, see voice agent architecture.

Why does TTS fall apart on long content?
The problem is architectural. Text and audio occupy fundamentally different token spaces. A second of speech carries 2-3 text tokens but needs 12.5-75 audio frames depending on the codec, often with several codebooks per frame. That roughly 4-40x imbalance means autoregressive TTS models generate far more output steps than they receive input tokens.
Each generation step introduces a small probability of error. Over a single sentence, these errors are negligible. Over thousands of steps (the territory of a five-minute clip), they compound. The model loses track of where it is in the text. Attention alignment between encoder and decoder degrades. Four failure modes appear:
Word skipping. The model advances past a word without generating audio for it. Common with repetitive text or complex phoneme sequences.
Repetition. A phrase is spoken twice when it appears once in the source. The attention mechanism circles back to a previous position.
Inserted words. Audio contains words absent from the source text. The model hallucinates content to fill perceived gaps.
Speaker drift. The synthesized voice gradually shifts toward a more generic timbre over extended passages. Chapter three of an audiobook sounds different from chapter one.
Most TTS models are evaluated on clips of a few seconds, and many are trained on utterances of seconds to tens of seconds. Quality often degrades once generation runs past the lengths seen in training. The gap between “works on a demo” and “works on a podcast episode” is enormous.
graph TD
A[Text: 2-3 tokens/sec] -->|4-40x expansion| B[Audio: 12.5-75 frames/sec]
B --> C[Autoregressive generation]
C --> D{Error at each step}
D -->|1 sentence| E[Negligible errors]
D -->|5 minutes| F[Compounding failures]
F --> G[Word skipping]
F --> H[Repetition]
F --> I[Hallucinated words]
F --> J[Speaker drift]
Production systems work around this with chunking. ElevenLabs segments text at sentence boundaries, generating each chunk separately and using previous/next context parameters to smooth prosodic transitions. Amazon Polly’s long-form engine processes up to 100,000 characters per request using internal chunking (exact segment size is not publicly documented). The stitching produces audible discontinuities at chunk boundaries: subtle enough for casual listening, obvious enough to derail an audiobook.
How does TADA avoid skipped and inserted words?
TADA (Text-Acoustic Dual Alignment), released by Hume AI on March 10, 2026 (paper: arXiv 2602.23068), attacks the token imbalance directly. Instead of generating dozens of audio tokens per text token and hoping the attention mechanism keeps them aligned, TADA enforces one-to-one synchronization. Each autoregressive step produces exactly one text token and one continuous acoustic vector. Text and speech advance in lockstep.
For skipping and insertion, this is a structural constraint rather than a learned behavior that sometimes fails. Every text token gets an acoustic representation, so the model cannot skip one, and there is no acoustic slot without a text token, so it cannot insert one. Hume reports zero transcription hallucinations (no skipped or invented words) across more than 1,000 test samples. The constraint does not guarantee that every word is pronounced well, and Hume notes that the voice can occasionally drift on longer texts.
The efficiency gain is large. At 25-50 audio frames per second, a 2,048-token window holds roughly 40-80 seconds of audio. At about 3 text tokens per second, the same window holds roughly 680 seconds, about 11 minutes of continuous speech; Hume quotes about 700 seconds of context. That is close to a 10x expansion from token efficiency alone.
The acoustic side uses continuous features rather than discrete tokens. The LLM produces a continuous acoustic vector per step, and a flow-matching stage (10 steps by default in Hume’s repo) generates the acoustic detail before decoding to audio. Continuous features avoid the information loss of discrete quantization.
TADA ships as a 1B English model and a 3B multilingual model, both built on Llama 3.2. The code is MIT-licensed; the weights fall under the Llama 3.2 Community License. Hume says TADA is more than five times faster than comparable LLM-based TTS systems, and its repo cites a real-time factor of about 0.12 on an H100 with cached prompts. For the cost economics of self-hosting models like this, see self-hosting TTS production economics.
The limitation: alignment fixes skipping and insertion. It does not by itself plan multi-speaker dialogue, emotional arcs, or scene-level context. For those, you need a different architectural philosophy.
What does Borderless do differently with chain-of-thought?
Borderless Long Speech Synthesis (arXiv 2603.19798, March 2026) reframes the problem. TADA asks: “How do I generate audio tokens that stay aligned with text tokens?” Borderless asks: “How do I plan what the audio should sound like before I generate it?”
The framework uses hierarchical annotation at three levels:
Global level. Scene semantics, overall emotional arc, speaker roster, acoustic environment. “This is a tense dialogue between two speakers in a quiet room, building toward a confrontation.”
Sentence level. Discourse context, prosodic boundaries, turn-taking patterns. “Speaker A interrupts here. Speaker B’s volume rises. This sentence ends with a question that expects no answer.”
Token level. Phonetic detail, fine-grained timing, emphasis placement. The lowest level of control.
This hierarchy serves as a structured semantic interface between an LLM agent and the synthesis engine. The agent uses chain-of-thought reasoning to plan the global and sentence levels before synthesis begins. Think of it as an audio director’s notes that the TTS engine follows.
The system uses a continuous tokenizer (not discrete) and supports capabilities that single-speaker systems cannot: multi-speaker synthesis with interruptions and overlapping speech, instruction-following synthesis (“make this part sound more urgent”), and VoiceDesigner for custom voice creation.
Where TADA prevents hallucination through alignment constraints, Borderless prevents incoherence through planning. The trade-off is clear: TADA is simpler and faster for single-speaker long-form content. Borderless handles complex multi-speaker scenarios that TADA’s architecture does not address.
How do these compare to production alternatives?
Two camps have formed in the long-form TTS space.
Continuous tokenizer camp (architectural solutions):
| System | Max duration | Speakers | Approach | Open |
|---|---|---|---|---|
| TADA (Hume AI) | ~11 min (Hume: ~700 s) | Not stated | 1:1 text-acoustic alignment | Yes (Llama license weights) |
| VibeVoice (Microsoft) | ~90 min | Up to 4 | 64K-token context, 7.5 Hz frame rate | Yes, but TTS code withdrawn Sept 2025 |
| Borderless | Extended | Multi, with overlaps | Agent-centric CoT planning | Release not stated |
Chunking camp (production workarounds):
| System | Max per request | Approach | Open |
|---|---|---|---|
| ElevenLabs | Per-request limit | Sentence-boundary chunking, context hints | No |
| Amazon Polly | 100,000 chars (async) | Internal chunking, segment size not documented | No |
| Fish Audio Story Studio | Extended | Chapter management, sentence-level revision | No (hosted product) |
Microsoft’s VibeVoice (1.5B and 7B parameters, August 2025) sits at the ambitious end of the continuous camp. It handles a 64K-token context, roughly 90 minutes of speech, with up to 4 speakers. The architecture uses a next-token diffusion framework with acoustic and semantic continuous tokenizers at 7.5 Hz, and it was designed for podcast and audiobook production. Microsoft removed the TTS code from its GitHub repository in September 2025 after misuse (see the VibeVoice deep dive).
Fish Audio’s Story Studio, launched in December 2025, takes a different approach. Rather than solving the length limit in the model, it builds tooling around it: chapter management, sentence-level regeneration, and multi-character voices kept consistent across chapters.
For audiobook production, where human narration commonly costs a few hundred dollars per finished hour, the quality bar is high and the economic incentive is real. ACX estimates about 9,300 words per finished hour, so a 50,000-word book runs about 5.5 hours of audio: at $200-400 per finished hour, that is roughly $1,100-2,200. AI generation costs a fraction, but only if quality survives the full duration.
Which approach fits which use case?
Single-speaker long content (reports, articles, accessibility): TADA. Ruling out skipped and inserted words matters when information fidelity is non-negotiable, and an 11-minute context handles most documents in a single pass.
Multi-speaker production (podcasts, dialogue, audiobooks with characters): Borderless or VibeVoice. Multi-speaker synthesis with interruptions, emotional arcs, and scene-level planning requires the scaffolding approach.
High-volume production pipelines: Fish Audio Story Studio or ElevenLabs. When you need revision control, chapter management, and human-in-the-loop editing, the tooling matters more than the model architecture.
Voice agents reading retrieved documents: TADA. A voice agent that reads a long policy document or research paper needs the content to be exactly right. Hallucinated words in a legal document are not acceptable.
Key takeaways
- A root cause is token imbalance. Codec-based audio needs roughly 4-40x more generation steps than text. Autoregressive errors compound over long sequences into skipping, repetition, and hallucination.
- TADA addresses it with alignment. One-to-one text-acoustic sync rules out skipped and inserted tokens by design, and fits about 11 minutes in a 2,048-token window. Voice drift on long texts is still possible.
- Borderless solves it with planning. Chain-of-thought reasoning plans prosody, emotion, and speaker interactions before generation begins.
- Production systems still chunk. ElevenLabs and Amazon Polly segment text and stitch audio. It works, with audible seams.
- VibeVoice pushes the boundary. 90 minutes, 4 speakers, 64K context: the most ambitious continuous approach in open research, though Microsoft withdrew the TTS code after misuse.
- The audiobook economics are real. $200-400/hr for human narration versus cents for AI. Quality on long content is the remaining barrier.
Further reading
- The sub-100ms TTS race: latency comparison of TADA, VoXtream2, Chatterbox, and API contenders
- TTS system fundamentals: architectural evolution from Tacotron to modern models
- Self-hosting TTS production economics: cost analysis at production scale