Add soniox provider

[dan2k3k4]

Relevant issues

Fixes #29507

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Screenshots / Proof of Fix

I have a proof-of-concept tool for testing transcription of audio into SRT and then into another language (if needed), so passing response_format=srt to the request should result in a subtitles format:

Type

🆕 New Feature
🐛 Bug Fix
🧹 Refactoring
📖 Documentation
🚄 Infrastructure
✅ Test

Changes

[dan2k3k4]

@greptileai

[codecov]

Codecov Report

❌ Patch coverage is 96.21381% with 17 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/llms/soniox/common_utils.py	93.18%	9 Missing ⚠️
litellm/llms/soniox/audio_transcription/handler.py	98.52%	3 Missing ⚠️
.../litellm_core_utils/get_supported_openai_params.py	33.33%	2 Missing ⚠️
litellm/main.py	50.00%	2 Missing ⚠️
.../llms/soniox/audio_transcription/transformation.py	98.96%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR adds Soniox as a new LiteLLM provider, supporting async speech-to-text transcription via Soniox's multi-step REST API (upload → create → poll → fetch → cleanup). The integration is well-structured: it follows established provider patterns, uses litellm's shared HTTP clients on both sync and async paths (including ssl_verify propagation), correctly passes model_response through to response building, and ships comprehensive mock-only unit tests.

• New handler (handler.py) orchestrates the polling pipeline with server-side-clamped poll intervals and best-effort cleanup; response transformation supports plain text, SRT, VTT, and verbose JSON formats synthesized from Soniox token timestamps.
• Provider is fully registered across LlmProviders enum, models_by_provider, lazy-import registry, model_prices_and_context_window.json, provider_endpoints_support.json, and the proxy UI credentials form.

Confidence Score: 5/5

New provider addition with no modifications to existing behaviour; safe to merge.

The change is purely additive — all existing code paths are untouched. The new Soniox handler correctly uses litellm's shared HTTP clients, propagates ssl_verify on both sync and async paths, and passes model_response through to response construction. Tests are mock-only and cover the full polling lifecycle. The only minor issue is a missing defensive guard on the poll_interval float conversion, which has no impact on correctness for valid inputs.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/soniox/audio_transcription/handler.py	New handler orchestrating Soniox's multi-step async transcription (upload → create → poll → fetch → cleanup) on both sync and async paths; both paths correctly use litellm's shared HTTP clients with ssl_verify propagation.
litellm/llms/soniox/audio_transcription/transformation.py	Transformation config mapping OpenAI params to Soniox-native params; handles language, response_format (srt/vtt/verbose_json), and webhook secret redaction correctly.
litellm/llms/soniox/common_utils.py	Shared utilities: SRT/VTT subtitle rendering from token timestamps, API key/base resolution, poll-interval clamping constants — all well-bounded.
litellm/main.py	Adds soniox dispatch branch in transcription() following the same pattern as nvidia_riva; type-ignore on provider_config is intentional and safe at runtime.
litellm/init.py	Registers soniox_models set, adds it to model_list and models_by_provider — consistent with all other provider registrations.
litellm/types/utils.py	Adds SONIOX to LlmProviders enum — single-line, correct placement.
model_prices_and_context_window.json	Adds soniox/stt-async-v4 model entry with pricing, mode, and supported endpoints; no unrelated modifications in this version of the diff.
tests/test_litellm/llms/soniox/audio_transcription/test_soniox_audio_transcription_handler.py	Comprehensive mock-based tests covering sync/async flows, upload, polling, cleanup, edge cases — no real network calls.
tests/test_litellm/llms/soniox/audio_transcription/test_soniox_audio_transcription_transformation.py	Tests transformation config including param mapping, response building, SRT/VTT rendering, and error handling — good coverage.
tests/test_litellm/llms/soniox/test_soniox_provider_registration.py	Tests provider enum registration, model list inclusion, lazy import, and get_llm_provider resolution — all mock-based.

_{Reviews (6): Last reviewed commit: "feat(soniox): add soniox audio transcrip..." | Re-trigger Greptile}

[greptile-apps]

Greptile Summary

This PR adds Soniox as a new audio transcription provider, implementing a multi-step async pipeline (file upload → create job → poll → fetch transcript → cleanup) behind LiteLLM's standard litellm.transcription() interface. The Soniox-specific code is well-structured, uses existing LiteLLM HTTP handlers, and is covered by comprehensive mock-only tests.

• The core Soniox provider files (handler.py, transformation.py, common_utils.py) are clean and follow established LiteLLM patterns; the litellm/main.py dispatch and provider registration hooks are correct.
• Both model_prices_and_context_window.json and its backup have unintended side-effects from JSON reformatting: \"supports_reasoning\": true was silently removed from bedrock/us-east-1/minimax.minimax-m2.5, bedrock/us-west-2/minimax.minimax-m2.5, and minimax.minimax-m2.5 (bedrock_converse), which would break any caller relying on reasoning for those models.
• provider_endpoints_support.json also lost the a2a and interactions fields from the charity_engine entry as a formatting side-effect.

Confidence Score: 3/5

Safe to merge for Soniox functionality, but the JSON reformatting removed supports_reasoning from existing bedrock minimax models, which is a regression for those users.

The Soniox provider implementation itself is solid. The concern is in the JSON files: supports_reasoning: true was dropped from three existing bedrock/minimax model entries as a side-effect of reformatting, and two endpoint fields were removed from charity_engine. These unrelated changes could silently break users who rely on reasoning with those models.

model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json — the minimax model entries need supports_reasoning restored; provider_endpoints_support.json needs the charity_engine a2a/interactions fields restored.

Important Files Changed

Filename	Overview
litellm/llms/soniox/audio_transcription/handler.py	New handler orchestrating Soniox's multi-step async transcription flow (upload → create → poll → fetch → cleanup) for both sync and async paths; uses existing LiteLLM HTTP handlers correctly and applies cleanup in finally blocks.
litellm/llms/soniox/audio_transcription/transformation.py	New transformation config mapping OpenAI params to Soniox equivalents and building TranscriptionResponse from Soniox payloads; logic is sound.
litellm/llms/soniox/common_utils.py	Shared constants, exception class, and utility helpers for Soniox; straightforward and correct.
litellm/main.py	Adds Soniox dispatch branch to transcription(); follows the same pattern as the nvidia_riva/elevenlabs branches.
model_prices_and_context_window.json	Adds soniox/stt-async-v4 entry, but JSON reformatting also silently removes supports_reasoning: true from bedrock/us-east-1/minimax.minimax-m2.5, bedrock/us-west-2/minimax.minimax-m2.5, and minimax.minimax-m2.5 (bedrock_converse) — a likely regression.
litellm/model_prices_and_context_window_backup.json	Mirrors the main pricing JSON changes, including the same unintended supports_reasoning removals for minimax models.
provider_endpoints_support.json	Adds soniox entry correctly, but also removes a2a and interactions fields from the charity_engine entry as a formatting side-effect.
litellm/llms/soniox/README.md	User-facing documentation added inside the litellm package; per contribution rules this should live in the litellm-docs repo instead.
litellm/init.py	Registers soniox_models set and adds it to model_list/models_by_provider; follows existing patterns correctly.
litellm/litellm_core_utils/get_llm_provider_logic.py	Adds soniox branch to resolve api_base and api_key from env vars; consistent with other provider branches.
tests/test_litellm/llms/soniox/audio_transcription/test_soniox_audio_transcription_handler.py	Comprehensive mock-only tests covering upload, poll, cleanup, error, and async flows; no real network calls.
tests/test_litellm/llms/soniox/audio_transcription/test_soniox_audio_transcription_transformation.py	Unit tests for the transformation config covering param mapping, response building, and token rendering.

_{Reviews (2): Last reviewed commit: "Merge branch 'BerriAI:litellm_internal_s..." | Re-trigger Greptile}

[dan2k3k4]

@greptileai

[veria-ai]

PR overview

All previously flagged issues have been addressed. No open security concerns remain on this pull request.

Security review

No open security issues remain on this pull request.

Fixed/addressed: 3 · PR risk: 0/10

[dan2k3k4]

@greptileai

[Sameerlite]

@greptileai

[Sameerlite]

The concern is in the JSON files: supports_reasoning: true was dropped from three existing bedrock/minimax model entries as a side-effect of reformatting, and two endpoint fields were removed from charity_engine. These unrelated changes could silently break users who rely on reasoning with those models.

Please fix this

[dan2k3k4]

The concern is in the JSON files: supports_reasoning: true was dropped from three existing bedrock/minimax model entries as a side-effect of reformatting, and two endpoint fields were removed from charity_engine. These unrelated changes could silently break users who rely on reasoning with those models.

Please fix this

Added a fix, squashed my commits into one and force pushed

[Sameerlite]

@greptileai