SEC 07 / MODELS & LATENCY

Choosing a model & latency

ElevenLabs exposes a models endpoint with capability information — supported operations, languages, character limits, rates, and other feature flags. Applications that need to support several models should inspect those capabilities rather than hardcoding every assumption; a model that gains a language or loses a limit shouldn’t require a code change to discover.

The three main TTS models

Model	Main strength	Language/length profile	Primary trade-off
`eleven_flash_v2_5`	Lowest latency	32 languages; up to 40,000 characters	Less robust number normalization by default
`eleven_multilingual_v2`	Quality and stability	29 languages; up to 10,000 characters	Higher latency and cost than Flash
`eleven_v3`	Expressiveness and multi-speaker output	More than 70 languages; up to 5,000 characters	Smaller request size and different latency profile

Per the models overview, Flash v2.5 runs at approximately 75 ms of model inference latency, excluding network and other end-to-end overhead. Multilingual v2 is positioned for higher-quality, stable multilingual output, while v3 is the expressive, multi-speaker option.

The Flash number-normalization trap

Flash v2.5 disables number normalization by default, as described in the models overview. That can affect dates, currencies, telephone numbers, abbreviations, and digit sequences. Multilingual v2 generally provides stronger normalization, and enterprise configurations can enable explicit text normalization where supported.

For important numerical content, preprocess it yourself. For example:

£1,250.40

may be transformed into:

one thousand two hundred and fifty pounds and forty pence

Similarly, telephone numbers may need deliberate grouping rather than relying on a model to infer the desired spoken form.

Where latency actually comes from

The latency optimization best practices emphasize five things:

selecting Flash for low-latency use cases;
beginning playback through streaming;
locating application servers near the ElevenLabs region;
using WebSockets only when text arrives incrementally;
using automatic generation-trigger behaviour for dynamic text streams.

Even after all five are applied, end-to-end latency still includes LLM generation time, network transit, buffering, decoding, and audio-device startup — model inference is only one segment of the pipeline a user actually experiences.

Output formats

Choose the output format for the downstream consumer, not merely for maximum theoretical quality.

Destination	Typical choice
Browser playback/download	MP3
Editing or further DSP	PCM/WAV-compatible raw audio
Low-latency application pipeline	PCM at the application’s native rate
Telephone network	`ulaw_8000` or the format required by the carrier
Archival	High-quality PCM or high-bitrate compressed audio

Avoid repeatedly transcoding. Every extra conversion adds latency, CPU usage, and potential quality degradation.

Chunk sizing for WebSocket TTS

With WebSocket TTS, very small text fragments can produce unnatural prosody and excessive overhead. Very large fragments delay the first generation. Use sentence- or clause-aware buffering rather than forwarding every LLM token individually.

For example:

Bad:
"The"
" total"
" is"
" 1"
","
"250"

Better:
"The total is one thousand two hundred and fifty pounds."