voice-latency

A per-hop latency budget for voice agents. It models two architectures, cascaded (STT then LLM then TTS) and realtime (one speech-to-speech model), adds framework overhead, and tells you where the milliseconds go.

The numbers are the point. Every provider figure lives in one Python file with a source and a date. If you have measured better, open a PR. That is how a latency benchmark stays honest.

This package powers the calculator at mahimai.ca/tools/latency-calculator; the site reads the same data, so the two never disagree.

Install

pip install voice-latency      # or: uv pip install voice-latency

Use it

from voice_latency import estimate

# Cascaded pipeline (default)
e = estimate(stt="Deepgram Nova-3", llm="GPT-4o mini", tts="Cartesia Sonic",
             transport="PSTN / SIP", endpointing=700, calls=25)
print(e.total_p50, e.total_p95)   # 1326 1589
print(e.max_hop)                  # 'endp'  (the endpointing delay you set)

# Realtime, speech-to-speech
r = estimate(mode="realtime", realtime="Gemini Live (Flash)",
             framework="LiveKit Agents", transport="WebRTC")
for hop in r.rows:
    print(f"{hop.label:<26} {hop.p50:>4}ms  p95 {hop.p95}ms")

From the terminal:

voice-latency                                   # default cascaded turn
voice-latency --mode realtime --realtime "OpenAI GPT Realtime"
voice-latency --framework "Google ADK (bidi streaming)" --transport WebRTC
voice-latency list                              # every provider it knows
voice-latency export --out data.json            # the dataset as JSON

The model

A turn's latency is the sum of independent hops. The p50 total sums the per-hop p50s. Percentiles do not add, so the p95 total adds each hop's spread (p95 minus p50) in quadrature on top of the p50 total. Hops:

  • cascaded: network in, jitter, STT, endpointing, LLM, TTS, network out
  • realtime: network in, jitter, endpointing, speech-to-speech model, network out
  • an orchestration framework, if any, adds one overhead hop

Constants (cross-region penalty, concurrency load, endpointing math, thresholds) live in MODEL in src/voice_latency/data.py.

The data

All of it is in src/voice_latency/data.py: STT, LLM, TTS, realtime speech-to-speech models, transport, and frameworks. Figures are illustrative defaults unless a source says otherwise. They are not vendor-published or independently benchmarked.

Improve them. See CONTRIBUTING.md. A PR that adds a source is worth more than a PR that just changes a number.

License

MIT.

Project links

Data verified by PyPI on Aug 21, 2026
Data provided by the project maintainers, verified at the time the release was uploaded to PyPI.

Key dates

PyPI data
Data sourced directly from PyPI's database.
  • Released:
Newer release available (0.1.1)

1 maintainer

PyPI data
Data sourced directly from PyPI's database.
Avatar for mahimairaja from gravatar.com mahimairaja

Credits

Author: Mahimai Raja J

GitHub Statistics

Data verified by PyPI on Aug 21, 2026
The GitHub source repository was provided by the project maintainers and verified by PyPI at the time of upload. Stars, forks, and open issues/PRs are derived from that repository and have not been independently verified.

License

MIT License (MIT)

Requires

Python >=3.9

Provides Extra

dev

Tags

latency livekit llm realtime speech-to-speech stt tts voice voice-agents

Classifiers

Report project as malware

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

voice_latency-0.1.0.tar.gz (10.5 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

voice_latency-0.1.0-py3-none-any.whl (10.3 kB view details)

Uploaded Python 3

File details

Details for the file voice_latency-0.1.0.tar.gz.

File metadata

  • Download URL: voice_latency-0.1.0.tar.gz
  • Upload date:
  • Size: 10.5 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for voice_latency-0.1.0.tar.gz
Algorithm Hash digest
SHA256 44fdf1767e41fdfc63039b9853fc732a4eb268636e4e170f0ccdbc2cb730793e
MD5 8b1bb2c030fdf8aa37cd3d12f6e18b55
BLAKE2b-256 f30d3294e5c2f7576c0fe6134702704997ac2371b06aea899dbed38cb3903218

See more details on using hashes here.

Provenance

The following attestation bundles were made for voice_latency-0.1.0.tar.gz:

Publisher: publish.yml on mahimailabs/voice-latency

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file voice_latency-0.1.0-py3-none-any.whl.

File metadata

  • Download URL: voice_latency-0.1.0-py3-none-any.whl
  • Upload date:
  • Size: 10.3 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for voice_latency-0.1.0-py3-none-any.whl
Algorithm Hash digest
SHA256 cc55fdb2f09b73312069ca72ed5d0f9d19e3bba8fdf81c2355eee698396638e4
MD5 3bad6ba6a7a7f4482dc4a39d4a8c6ced
BLAKE2b-256 aa7b6911f9d067fc9bd3bcc719228365c71046202b982d771df933c8095ac762

See more details on using hashes here.

Provenance

The following attestation bundles were made for voice_latency-0.1.0-py3-none-any.whl:

Publisher: publish.yml on mahimailabs/voice-latency

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Release history Release notifications | RSS feed

0.1.1

2 files

This release

0.1.0 This release

2 files

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page