# Transcribe a live caller for a voice agent, with a clear end of turn

> Transcribe a live caller for a voice agent, with a clear end of turn: 6 companies fully do it, 13 with a limit, 50 carry the tag. Every listing in Speech-to-text with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/speech-to-text.voice-agent-live/
- Markdown: https://www.anchorterminal.com/niches/speech-to-text.voice-agent-live/index.md (~2,300 tokens)
- Slim: https://www.anchorterminal.com/niches/speech-to-text.voice-agent-live/index.min.md (~480 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/speech-to-text.voice-agent-live/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 6 | 13 | 50 | 18 | 6 |

## What counts as done

A final transcript event for each turn, with an explicit end-of-turn or finalisation signal, in a session that can last a 10-minute call.

Measured, for our own tests: the final event arrives within 500 ms of end of speech at the median, and no session silently stops returning text over a 10-minute call.

How agents ask for it: transcribe what the caller is saying as they speak; give me the final text when they stop talking; wire speech-to-text into my phone agent

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **AssemblyAI Speech-to-Text (Universal) (AssemblyAI, Inc.)**: Turn messages carry end_of_turn true; Terminate returns in-flight turns then Termination. Pipecat measured 282 to 307 ms median to final. acc-31's stall was on the retired v2 transcriber.
- **Cartesia Ink (Cartesia AI, Inc.)**: Ink 2 emits turn.start, turn.eager_end, turn.resume and turn.end; close flushes remaining audio and returns a done acknowledgment. Pipecat measured 299 ms median, 1,584 ms at p99.
- **ElevenLabs Scribe Speech to Text API (Eleven Labs Inc.)**: committed_transcript is the final per segment; commit_strategy turn_prediction on scribe_v3_live commits when the model predicts the turn ended. Pipecat measured scribe_v2_realtime at 281 ms median.
- **Inworld STT (Theai, Inc. dba Inworld AI)**: Finals carry isFinal true per turn with endOfTurnConfidenceThreshold and client endTurn; closeStream finalises pending audio. No independent latency row, and headers rather than query carry the key.
- **Soniox Speech-to-Text (Soniox Inc.)**: enable_endpoint_detection finalises the segment and returns an <end> token; finalize forces a final. Pipecat measured stt-rt-v6 at 266 ms median; streams cap at 300 minutes.
- **Speechmatics Speech-to-Text (Cantab Research Ltd)**: Agent STT emits StartOfTurn and EndOfTurn, with AddSegment finals and EndOfStream then EndOfTranscript on close; Pipecat measured linden-1 at 369 ms median. Temporary keys travel in the WebSocket URL.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Amazon Transcribe (Amazon Web Services, Inc.)**: StartStreamTranscription returns segments with IsPartial false, but no end-of-turn event, and the Pipecat benchmark measured 1,136 ms median time to final, above the 500 ms target.
- **Azure AI Speech speech-to-text (Microsoft Corporation)**: Speech SDK recognized events are per-segment finals with no turn event; MAI-Transcribe-2-Streaming has no server VAD and needs a client commit. Pipecat measured 1,016 ms median to final.
- **Deepgram Speech-to-Text (Nova-3, Flux) (Deepgram, Inc.)**: Flux emits EndOfTurn with eot_threshold and eot_timeout_ms; Nova-3 streaming gives speech_final and CloseStream flushes. Pipecat measured Nova-3 at 247 ms, but acc-28 reports sessions silently stopping with no error.
- **Giggy Speech to Text (Giggy)**: The live WebSocket returns cumulative transcription events with is_final and input_complete, but the docs say turn boundaries and VAD are the caller's job; no end-of-turn event and no independent latency row.
- **Gladia Speech-to-Text API + MCP (Gladia SAS)**: Live init has endpointing (silence seconds) and maximum_duration_without_endpointing, and receive_final_transcripts sends finals; no end-of-turn event beyond the final, and Gladia has no Pipecat latency row.
- **Google Cloud Speech-to-Text (Google LLC)**: StreamingRecognize has SPEECH_ACTIVITY_END voice events and is_final results, but streams stop at 5 minutes and must be reopened; Pipecat measured latest-long at 878 ms and gemini-3.5-transcribe-live at 458 ms.
- **IBM Watson Speech to Text (International Business Machines Corporation)**: The WebSocket returns final results per phrase with interim_results and end_of_phrase_silence_time, no explicit end-of-turn event, and the access token travels in the URL query string; no independent latency row.
- **Mistral Voxtral Transcribe (Mistral AI (RCS Paris 952 418 325))**: Realtime 2602 streams text deltas and a transcription.done event at the end, with no end-of-turn event; Pipecat measured it at 525 ms median. Realtime 3 (Public Preview) claims end-of-turn detection.
- **Nemotron 3.5 ASR Streaming 0.6B (NVIDIA Corporation)**: Cache-aware streaming at 80 to 1,120 ms chunks; Pipecat measured 236 ms median to final, but the model emits text, not an end-of-turn event, so the caller supplies VAD and endpointing.
- **OpenAI Speech to Text (OpenAI OpCo, LLC)**: gpt-live-transcribe needs the client to commit each turn (no server_vad), then sends completed; acc-21 saw deltas arrive with the final. Pipecat measured gpt-4o-transcribe at 637 ms median.
- **Qwen3-ASR-0.6B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: Streaming runs only through vLLM's realtime path, without timestamps or batching, and emits text rather than an end-of-turn event; no independent latency row.
- **Qwen3-ASR-1.7B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: Streaming runs only through vLLM's realtime path, without timestamps or batching, and emits text rather than an end-of-turn event; no independent latency row.
- **Rev AI Speech-to-Text API (Rev.com, Inc.)**: Streaming returns partial and final hypotheses with no end-of-turn event, the access token travels in the URL query string, and Rev AI has no independent latency row.
- **Smallest.ai Pulse STT (Smallest, Inc.)**: Live WebSocket has endpointing with eou_timeout_ms, is_final transcripts, speech_ended events and finalize and close_stream messages; Pipecat measured pulse at 398 ms median but 1,593 ms at p99, and keys may travel in the query string.
- **Voxtral Mini 4B Realtime 2602 (Mistral AI (RCS Paris 952 418 325))**: vLLM serves it on /v1/realtime with a 480 ms delay, but it emits text only, no end-of-turn event; Pipecat measured the hosted twin at 525 ms median and 1,913 ms at p99.

- not this job: Cohere Transcribe 03-2026 (Cohere Inc.): Batch model with no streaming runtime or endpoint; Cohere's docs list audio-in text-out only.
- not this job: Groq Speech-to-Text (Groq LLC): No streaming or realtime endpoint; files or URLs only.
- not this job: Ming-3.0-asr-tiny (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI)): Batch model; no streaming runtime, only SGLang-Omni server-sent events of a finished transcript.
- not this job: Ming-3.0-audio-flash (inclusionAI, Ant Group (named in the inclusion-ai.org footer; no registered company named)): Offline audio-language model; no streaming transcription path is documented.
- not this job: Ming-3.0-audio-tiny (inclusionAI (Ant Group)): Offline audio-language model; no streaming transcription documented.
- not this job: MiniMax Speech-to-Text (Nanonoble Pte. Ltd.): Files only; stream=true streams text output of an uploaded file, with no live audio input.
- not this job: Parakeet TDT 0.6B v3 (NVIDIA Corporation): An offline TDT model with no streaming runtime or end-of-turn signal; the speech.streaming tag is not supported by the card.
- not this job: Underdog Bark 3.0 (Conway Research): Offline model run with Conway's loader; no streaming path.

3 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- no streaming endpoint (Groq, Cohere Transcribe, Ming, Parakeet TDT without a streaming runtime)
- stream drops after 5 minutes without reconnect (Google 5-minute stream limit)
- session stops returning text with no error, as in acc-28 and acc-31
- credentials in the WebSocket query string where logs can keep them

## Around it

Before: telephony.answer-call, voice.capture-microphone. After: llm.reply, text-to-speech.speak.

## Where this comes from

Weight 18 of 100 in Speech-to-text: 6 of 32 accounts (acc-04, acc-21, acc-28, acc-29, acc-31, acc-11) plus two independent benchmarks built for this job alone; the use case vendors lead with

As data: https://www.anchorterminal.com/api/v1/use-cases/speech-to-text.voice-agent-live.json, or the MCP tool get_use_case with id speech-to-text.voice-agent-live. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
