Gemini 3.5 vs Soniox v5 vs Linden-1: STT Guide 2026

Three new or recently updated speech-to-text options now compete for production voice-agent workloads. Google made Gemini 3.5 Transcribe and Transcribe Live generally available on August 26, 2026, Soniox released stt-rt-v5 on June 16, and LiveKit Agents added Gemini 3.5 Transcribe Live and Speechmatics Linden-1 on August 27 (Google changelog, Soniox model documentation, LiveKit Agents releases).

This is a research-based comparison, not a lab benchmark. No neutral, reproducible test currently ranks these three exact models on the same German 8 kHz telephone dataset. The useful question is therefore not "Which model wins overall?" It is which access path, session limit, telephone format, endpointing behavior, language handling, data route, and billing model fit your calls. If you need the basics first, start with what speech-to-text means.

Key Takeaways

• Gemini offers broad direct multilingual coverage and native LiveKit Inference access, but live sessions are limited to ten minutes and the LiveKit route cannot be regionally pinned.
• Soniox combines long sessions, telephone-native codecs, live diarization, code-switching, and configurable semantic endpointing through a plugin that uses your own Soniox key.
• Linden-1 is a native LiveKit Inference option with server-side end-of-speech detection, but its local vad input is ignored and no Linden-specific regional guarantee is published.
• Public prices describe different access paths and billing units. They are not a quality ranking.
• A decision should follow a controlled A/B test on your own calls, especially German public telephone network (PSTN) audio, entity capture, turn boundaries, and actual cost per conversation minute.

Quick comparison

The table is a snapshot dated August 29, 2026. It deliberately avoids accuracy scores because vendor benchmarks do not use a shared dataset, telephone channel, language mix, or endpointing setup.

Access and integration paths

The three names hide different deployment models that affect credentials, routing, and invoicing. A broader checklist for choosing an STT provider helps separate provider-level requirements from model-level tests.

Gemini 3.5 Transcribe Live

The direct model ID is gemini-3.5-transcribe-live; LiveKit exposes it as google/gemini-3.5-transcribe-live. It supports interim and final results over a live WebSocket session. Google's direct documentation also describes strict transcription and Smart Transcription behavior, plus automatic, hybrid, and manual voice-activity strategies (Gemini model page, live transcription guide).

The important operational constraint is the ten-minute session maximum. Long calls need a planned rollover, overlap, and transcript reconciliation strategy. Gemini live transcription also lacks word timestamps and live speaker diarization. The LiveKit Gemini page currently exposes a narrower language-code list than Google's direct documentation and does not establish that every direct VAD control is configurable through LiveKit Inference.

Soniox stt-rt-v5

Soniox uses the LiveKit plugin path rather than LiveKit Inference, so the application supplies its own Soniox key. Set stt-rt-v5 explicitly. Current Node.js plugin documentation may still show v4 as a default, while Soniox aliases v4 to v5 server-side. The Soniox pipeline update provides related context.

Soniox documents real-time streaming sessions of up to 300 minutes (real-time transcription, limits and quotas). This is useful for long consultations, but teams still need reconnect and failure handling.

Speechmatics Linden-1

Linden-1 is accessed as speechmatics/linden-1 through LiveKit Inference, which handles model routing and usage billing. A separate public direct-API SKU for this exact model was not verified. Speechmatics positions Linden for voice agents, and LiveKit recommends it for most Speechmatics agent workloads (Speechmatics announcement, LiveKit Speechmatics page).

End-of-speech is detected server-side. A local LiveKit vad object is ignored for Linden, so teams should not assume that swapping local VAD libraries will change its turn boundary.

German and multilingual conversations

All three options include German, but their published language claims are not directly equivalent. Google states 85+ languages for the direct model, while LiveKit currently lists 26 language codes for its Gemini integration. Gemini can change languages during a session. Soniox documents 60+ languages and automatic code-switching, including switches within a sentence (Google live guide, Soniox supported languages). Linden's LiveKit page lists 55+ languages, but the reviewed sources do not clearly confirm arbitrary automatic code-switching.

For German production traffic, test names, compound nouns, dialects, accented German, and German-English switches. Add domain terms with Gemini custom vocabulary, Soniox context and terms, or Speechmatics additional_vocab. Test vocabulary against real utterances using these guidelines for specialist terms and names.

Turn detection matters beyond transcript accuracy

Voice agents act when the STT system decides a user has finished. A correct transcript that arrives after a long pause can still create a poor conversation, while an early endpoint can cut off a date, address, or correction. A current Hacker News discussion is useful only as a practitioner signal that turn-taking deserves production testing. Its comments and reported latency figures are not a controlled comparison of these models.

Google documents automatic server VAD, a hybrid strategy that combines server speech-start detection with the local audio_stream_end signal, and fully manual activity markers. The direct hybrid mode can avoid waiting for the server timeout, but configuration parity through LiveKit should be checked rather than assumed.

Soniox explicitly describes semantic endpointing based on pauses, intonation, speech patterns, and context. It exposes endpoint_latency_adjustment_level from 0 to 3, endpoint_sensitivity from -1.0 to 1.0, and max_endpoint_delay_ms from 500 to 3,000 ms. More aggressive settings can increase segmentation and slightly affect word and diarization accuracy, according to the Soniox endpoint detection documentation.

Linden detects the speech end on the server. Do not call this semantic endpointing without a Linden-specific primary source. The architectural consequences are easier to assess alongside the pipeline and realtime architecture guide and the voice-agent latency breakdown.

Telephone audio and structured data

Soniox documents direct carrier-audio support because its real-time API accepts PCM, mu-law, A-law, and other formats. Gemini's direct live-transcription specification calls for 16-bit PCM at 16 kHz mono, which can require resampling from an 8 kHz PSTN stream. LiveKit abstracts media transport for Linden, but that does not prove a Linden-specific raw-codec capability.

Score structured fields separately: names, spelled email addresses, telephone numbers, IBANs, appointment times, postal codes, and product codes. Measure entity error rate as well as word error rate. Gemini has no live diarization, Soniox does, and Speechmatics diarization should be validated on the exact Linden route.

Privacy and processing regions

No single EU-hosting statement covers all three paths. LiveKit Inference is zero data retention by default, but Gemini 3.5 Transcribe Live runs through one global deployment and cannot be selected for a region-pinned request. That prevents a strict regional-only claim for this route.

Soniox documents dedicated EU domains, including stt-rt.eu.soniox.com, regional processing, and regional storage. EU data residency must be enabled first (Soniox data residency). No Linden-specific region guarantee through LiveKit was verified. Confirm the route contractually before an EU-only rollout and review other voice AI STT deployment choices.

Price comparison by access route

The Soniox estimate follows its token pricing and reference consumption. Gemini's estimate follows Google's Gemini API pricing. The two LiveKit figures come from LiveKit Inference pricing. Compare invoices only after normalizing what each route includes. Lower STT cost does not imply lower total cost or higher recognition quality.

A neutral A/B test for real calls

Use identical audio, downstream logic, and measurement windows. Include:

1. German 8 kHz mu-law PSTN recordings, plus the resampled version required by an integration.
2. Standard German, regional dialects, non-native speakers, and German-English code-switching.
3. Names, email addresses, IBANs, telephone numbers, times, dates, and product codes.
4. Silence, thinking pauses, self-corrections, overlapping speech, background noise, and barge-in.
5. Short service calls and calls longer than ten minutes, including a controlled Gemini session rollover.

Report word and entity error rates, time from the final speech frame to a stable transcript, early and late endpoints, partial churn, barge-in errors, reconnects, and actual cost per conversation minute. Review by scenario because one average can hide failures on business-critical fields.

Decision matrix

This matrix prioritizes a test order, not a universal winner. Also verify whether each exact model is available in your current workspace before rollout.

Frequently asked questions

Which model is most accurate for German telephone calls?

There is no neutral head-to-head benchmark for these exact versions on the same German 8 kHz telephone dataset. Test your own calls and report both word error rate and entity error rate. Vendor scores from different datasets cannot establish a fair ranking.

Can Gemini 3.5 Transcribe Live handle calls longer than ten minutes?

Not as one uninterrupted documented session. The published maximum is ten minutes, so longer calls require a rollover design with overlap, transcript reconciliation, and failure testing.

Is Soniox stt-rt-v5 automatically processed in the EU?

No. Soniox offers dedicated EU endpoints and regional storage, but data-residency access must be enabled. Confirm the endpoint, storage behavior, and contract before production.

Does Linden-1 use semantic endpointing?

The reviewed primary sources confirm server-side end-of-speech detection and that local LiveKit VAD is ignored. They do not confirm semantic endpointing for Linden-1, so that label should not be used without newer model-specific documentation.

Conclusion and next step

Gemini 3.5 Transcribe Live, Soniox stt-rt-v5, and Speechmatics Linden-1 solve different deployment problems. Gemini emphasizes direct multilingual reach and native LiveKit access, Soniox exposes telephone and endpointing controls through a provider plugin, and Linden offers a LiveKit-managed route with server-side turn completion. None of those differences proves an overall accuracy winner.

Choose a shortlist from the decision matrix, verify model availability in your workspace, and run the same German call set through every candidate. If you want to evaluate that configuration against real conversations, book a personal Famulor demo to review your call flow, languages, and telephony environment together.

About the author

Sarah Müller is the named author of this Famulor comparison. The analysis uses public primary documentation and separates published facts from estimates and practitioner signals. For editorial or product questions, contact Famulor Support. Famulor's handling of personal data is described in its privacy information.