Speech-to-Text API Pricing & Speaker Diarization Compared (2026)
Live November 2026 per-minute and per-hour rates across 8 transcription APIs—covering batch vs. streaming, diarization add-on fees, multichannel multipliers, OpenAI's 2027 Whisper cutoff, and raw STT vs. all-in-one media pipelines.
Michael NguyenTested November 2026
Try Transcribe audio from Audo
Quick answer
In November 2026, raw speech-to-text APIs like AssemblyAI ($0.15–$0.21/hr batch, +$0.02/hr diarization), Google Cloud V2 Dynamic Batch ($0.18/hr), ElevenLabs Scribe v3 ($0.20/hr), Gladia (€0.16–€0.27/hr), and Deepgram Nova-3 ($0.26–$0.31/hr) hold the lowest per-hour floor for bulk text-only jobs or sub-300 ms WebSocket streaming, while Audo AI ($0.03/min or $1.80/hr list, $1.13–$1.44/hr on non-expiring $10–$100 packs from $5, zero monthly minimum) wins end-to-end audio and video pipelines in one API. Audo AI pairs 60-language batch transcription and automatic 8-speaker diarization ($0 extra) with 10 GB / 10-hour raw video uploads, keyless curl testing (POST /api/try/transcribe, 1 min free 3x/day), custom volume rates on /developers, a hosted MCP server, built-in noise removal ($0.05/min), and $0.01/min transcript_file_id chaining into filler cutting, bleeping, highlights, show notes, and burned-in captions.
| Provider & models | Batch STT rate | Streaming rate | Speaker diarization | Limits, free tier & billing traps |
|---|---|---|---|---|
| Audo AI (transcribe_audio) | $0.03/min ($1.80/hr) list; $1.13–$1.44/hr on $10–$100 packs; volume tiers at /developers | Batch / near-real-time only (<10 s per 2-min clip; <2 min per 1-hr file) | Included ($0 extra), on by default, up to 8 speakers, 60 languages; JSON, TXT, SRT, VTT | 10 GB / 10 hr raw audio or video; no multichannel multiplier; keyless 1 min free 3x/day; REST + hosted MCP + $0.01/min media chaining |
| Deepgram (Nova-3 / Flux) | $0.0043/min ($0.26/hr) mono; $0.0052/min ($0.31/hr) multi; Whisper Large $0.29/hr | $0.0048/min ($0.29/hr) mono; $0.0058/min ($0.35/hr) multi (promo); Flux $0.39–$0.47/hr | Included ($0) pre-recorded; +$0.0020/min (+$0.12/hr) on live streaming | 2 GB cap, 10-min 504 timeout (extract audio from large video); $200 free starter balance (no card); multichannel bills per channel; stateless |
| AssemblyAI (STT-3.7 / Universal-2) | $0.0035/min ($0.21/hr) STT-3.7 (32 langs); $0.0025/min ($0.15/hr) Universal-2 (99 langs) | $0.005/min ($0.30/hr) Realtime; $0.0075/min ($0.45/hr) Realtime-Agent (session time) | +$0.02/hr async ($0.17–$0.23/hr total, <=30 speakers); +$0.12/hr streaming inline | 2.2 GB direct upload (5 GB URL), 10 hr max; $50 free signup balance (no card); streaming bills open idle sockets; multichannel bills per channel |
| OpenAI Audio API (gpt-transcribe) | $0.0045/min ($0.27/hr) gpt-transcribe; deprecated whisper-1 & gpt-4o-transcribe $0.36/hr ($0.18/hr mini) | $0.017/min ($1.02/hr) on gpt-live-transcribe and gpt-realtime-whisper | Deprecated: legacy gpt-4o-transcribe-diarize ($0.36/hr) shuts down Feb 26, 2027; none on gpt-transcribe or Realtime | Hard 25 MB upload wall (requires local ffmpeg splitting); no free API tier; gpt-transcribe omits word timestamps, SRT, and VTT |
| ElevenLabs Scribe API (v3 / v2) | $0.0033/min ($0.20/hr) Scribe v3, v2 & v2 Medical across 90+ languages | $0.0037/min ($0.22/hr) Scribe v3 Live & v2 Realtime (~150 ms latency) | Included ($0 extra): <=32 speakers batch; inline live diarization on Scribe v3 Live | 5 GB / 15 hr audio & video; silence counts in billed duration; keyterms +$0.05/hr, editing +$0.07/hr; separate Voice Isolator is $7.20/hr |
| Gladia (Solaria-3 / Solaria-1) | €0.16/hr (€0.0027/min) Solaria-3 (5 EU langs); €0.27/hr (€0.0045/min) Solaria-1 (100+ langs) | €0.68/hr (€0.0113/min) live on Solaria-1 (3-hr max session) | Included (€0 extra) on async pre-recorded jobs only; unavailable on live streams | 1,000 MB / 135 min async cap (4h15 Enterprise); €50 starter balance (1 yr); automatic volume tiers to €0.12/hr; self-serve trains unless opted out |
| Google Cloud Speech-to-Text V2 | $0.003/min ($0.18/hr) Dynamic Batch; $0.016/min ($0.96/hr) Standard (drops at 500k+ min) | $0.016/min ($0.96/hr) Standard streaming (~5-min max per stream) | Included ($0 extra) on Chirp 3 Batch/Sync (14 locales); unsupported on Chirp 2 | 10 MB / 1 min inline cap; longer jobs need Google Cloud Storage (billed extra; Chirp 3 caps at 1 hr, or 20 min with word timestamps); multichannel bills per channel |
| Amazon Transcribe (US East) | $0.006/min ($0.36/hr) Standard Batch, billed per second ($0.0001/s) | $0.010/min ($0.60/hr) Standard Streaming ($0.0001667/s) | Included ($0 extra) on batch & streaming, up to 30 speakers | 60 min/mo free for 12 months; 2 GB / 8 hr cap; batch requires Amazon S3 staging; 2 channels included (no stereo surcharge); PII redaction +$0.144/hr |
Listen and read along
Press play to see who says what.
Transcript
Voices in these examples are AI-generated.
Four billing traps behind headline speech-to-text API rates
Headline speech-to-text rate cards ($0.15 to $0.36 per hour) assume clean, mono, pre-chunked audio files and zero downstream processing. Production engineering bills routinely land 2x to 4x higher once real user uploads hit four contract rules:
Multichannel stereo multipliers (2x on Deepgram, AssemblyAI, and Google V2)
Deepgram, AssemblyAI, and Google Cloud Speech-to-Text V2 bill each audio channel independently when multichannel processing is active—turning a 1-hour dual-channel call into 2 billable hours ($0.42/hr on AssemblyAI STT-3.7, $0.52/hr on Deepgram Nova-3 mono). Amazon Transcribe includes up to 2 channels at one duration rate, and Audo AI applies zero channel multiplier.
Open WebSocket idle billing and silence metering
AssemblyAI bills streaming endpoints by open WebSocket session duration rather than spoken audio—an idle socket left connected for 20 minutes between calls bills 20 full minutes plus any active streaming add-ons. ElevenLabs Scribe likewise meters every second of silence in both batch and live requests, while Google Cloud caps each streaming session around 5 minutes.
OpenAI's February 26, 2027 deprecation deadline and 25 MB upload wall
OpenAI deprecated whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize with shutdown scheduled for February 26, 2027. Its replacement, gpt-transcribe ($0.0045/min or $0.27/hr), drops speaker diarization, word-level timestamps, and native SRT/VTT outputs—and every /v1/audio/transcriptions call still hits a hard 25 MB ceiling that forces client-side ffmpeg splitting.
Cloud bucket staging and large-video timeouts
Amazon Transcribe rejects direct HTTP batch uploads until you stage files in Amazon S3 (capped at 2 GB and 8 hours). Google Cloud V2 rejects local files above 10 MB or 1 minute unless staged in Google Cloud Storage, and Deepgram enforces a 2 GB payload limit and 10-minute 504 timeout while advising teams to demux audio out of large MP4 files first.
Speaker diarization API pricing: batch vs. real-time
Engineers evaluating a speaker diarization api need to separate batch diarization (usually bundled or a small surcharge) from inline streaming diarization (where prices jump sharply or disappear altogether). Even among batch engines, maximum speaker counts and subtitle formatting differ widely.
For pre-recorded multi-speaker audio, ElevenLabs Scribe v3 ($0.20/hr, up to 32 speakers), Gladia Solaria-3/1 (€0.16–€0.27/hr), Speechmatics Melia 1 ($0.24/hr pre-recorded, $0.40/hr real-time with $100 starter balance), Deepgram Nova-3 ($0.26–$0.31/hr), and Amazon Transcribe ($0.36/hr, up to 30 speakers) include diarization at $0 extra. AssemblyAI adds $0.02/hr ($0.17/hr on Universal-2, $0.23/hr on STT-3.7, up to 30 speakers on recordings over 10 minutes). Google Cloud V2 includes diarization on Chirp 3 across 14 locales ($0.18/hr Dynamic Batch, $0.96/hr Standard), though Chirp 3 caps batch jobs at 1 hour—or 20 minutes when word timestamps are enabled—and Chirp 2 omits diarization completely.
Streaming diarization adds $0.12/hr—or drops out entirely
Turning on live diarization adds +$0.0020/min (+$0.12/hr) on Deepgram Nova-3 ($0.41–$0.47/hr promo) and +$0.12/hr on AssemblyAI STT-3.7-Realtime ($0.42/hr) and Realtime-Agent ($0.57/hr). ElevenLabs Scribe v3 Live bundles inline diarization at $0.22/hr (incompatible with its background audio filter), Amazon Transcribe includes it at $0.60/hr, and neither Gladia Live nor OpenAI Realtime supports speaker diarization.
Audo AI speaker diarization: $0 add-on, clean turn boundaries, 4 outputs in one job
On Audo AI, speaker_labels defaults to true at zero extra per-minute cost across 60 languages on files up to 10 hours and 10 GB, labeling up to 8 speakers (Speaker 1 through Speaker 8) with millisecond word timings. Every completed job returns JSON, speaker-turn TXT, SRT, and WebVTT (<v Speaker N>) together without writing custom subtitle formatters, and degrades gracefully to a full word-timed transcript if degenerate audio prevents speaker clustering.
Raw STT API vs. an all-in-one audio and video API
Comparing Audo AI ($1.80/hr list, $1.13–$1.44/hr on $10–$100 non-expiring packs) against a $0.20/hr raw speech-to-text endpoint only makes sense when text JSON is your final output. The moment a SaaS app, content pipeline, or AI agent also needs noisy-mic cleanup, MP4 video pass-through, filler cutting, spoken PII bleeping, highlight clipping, or burned-in word captions, raw STT is only step one of a multi-vendor engineering project.
On Audo AI, uploading a raw MP4, MOV, MKV, WebM, or AVI file once (up to 10 GB via multipart upload) unlocks 49 tools on the same file IDs—including keyless prototyping at POST https://audo.ai/api/try/transcribe (first minute free, 3x/day, no API key), exact script synchronization via align_script ($0.03/min, first 2 minutes free 3x/day on /text-to-srt), and the remote MCP server at https://audo.ai/mcp. Reusing the returned transcript_file_id across seven downstream tools drops their rate to $0.01–$0.02/min ($0.60–$1.20/hr), with custom volume discounts available within one day on /developers#volume (no sales call, no annual contract).
One transcript_file_id feeds 7 downstream tools at $0.01–$0.02/min
After one transcribe_audio run ($0.03/min), passing transcript_file_id cuts Auto captions (add_captions), Remove filler words (remove_fillers), Bleep censor (censor_audio, +$0.01/min for automatic AI PII with $0.03 paid-run minimum, +$0.05/min to preserve background music), AI clip finder (find_highlights, $0.03 min), Podcast show notes generator (write_show_notes, $0.03 min), and Meeting shortener (shorten_transcript) to $0.01/min ($0.60/hr), and Audio translator (translate_transcript) to $0.02/min ($1.20/hr).
Built-in denoising and bit-for-bit MP4 stream-copy before transcription
When source audio has HVAC hum, wind, or room reverb, calling Audo AI's Remove noise or Audio enhancer ($0.05/min list, $1.88–$2.40/hr on packs) cleans speech and normalizes to -16 LUFS while stream-copying H.264/HEVC video frames untouched—whereas ElevenLabs charges $0.12/min ($7.20/hr) for audio-only Voice Isolator capped at 500 MB / 1 hour.
| Pipeline step (1-hr stereo MP4) | Raw STT-only API + DIY stack | Audo AI single REST API / MCP |
|---|---|---|
| 1. Ingest & audio extraction | Self-hosted S3/GCS staging + ffmpeg demux/downmix worker (required to avoid 25 MB–2 GB caps & 2x stereo billing) | $0 extra — direct multipart upload up to 10 GB / 10 hr; zero multichannel multiplier |
| 2. Speech cleanup & -16 LUFS | Separate vendor API (e.g. ElevenLabs Voice Isolator $7.20/hr audio-only, 1-hr cap) or self-hosted CUDA GPU + ffmpeg remux | $3.00/hr list ($1.88–$2.40/hr on packs) via remove_noise or enhance_voice; video copied bit-for-bit |
| 3. Diarized transcript + SRT/VTT | $0.17–$0.36/hr raw JSON (AssemblyAI/Deepgram/ElevenLabs/AWS) + custom SRT/VTT speaker formatter | $1.80/hr list ($1.13–$1.44/hr on packs) via transcribe_audio; 8 speakers, JSON + TXT + SRT + VTT |
| 4. Cut filler words + show notes + highlights | Separate LLM token billing + custom frame-accurate ffmpeg video slice & A/V sync pipeline | $0.60/hr each ($1.80 total list; $1.13–$1.44 on packs) reusing transcript_file_id across remove_fillers, write_show_notes, and find_highlights |
| 5. Styled word-by-word burned captions | Custom Remotion / FFmpeg ASS renderer + emoji/font/GPU render fleet | $0.60/hr list ($0.38–$0.48 on packs) via add_captions reusing transcript_file_id (36 presets) |
When Deepgram, AssemblyAI, ElevenLabs, or AWS/GCP beats Audo AI
Audo AI is priced for complete audio/video processing pipelines rather than bare-metal transcription arbitrage. Choose a dedicated speech-to-text API instead when your workload fits one of these four profiles:
Sub-300 ms live WebSocket streaming (voice agents, telephony, live captions)
Audo AI processes uploaded files and URLs rather than open real-time audio sockets. For live voice agents or call-center assist, use Deepgram Nova-3/Flux ($0.29–$0.47/hr streaming), AssemblyAI STT-3.7-Realtime or Realtime-Agent ($0.30–$0.45/hr), or ElevenLabs Scribe v3 Live ($0.22/hr).
High-volume raw text-only transcription (500+ hours/month, zero media editing)
If you ingest thousands of call recordings strictly for CRM search, sentiment analytics, or LLM summaries and never denoise, slice, redact, or caption media files, AssemblyAI Universal-2 ($0.17/hr diarized), Google Cloud V2 Dynamic Batch ($0.18/hr), ElevenLabs Scribe v3 ($0.20/hr), Gladia Solaria-3 (€0.12–€0.16/hr), or Deepgram Nova-3 ($0.26/hr) will beat Audo AI's list pricing by 5x to 10x.
Instant self-serve HIPAA BAAs or native S3/GCS enterprise governance
Teams requiring a signed HIPAA BAA out of the box (AssemblyAI self-serve, Deepgram/Speechmatics Enterprise, AWS Transcribe, Google Cloud STT V2) or strict VPC-locked AWS S3 / GCP Cloud Storage data paths should stay on those enterprise speech stacks.
More than 8 speakers per recording or offline zero-egress compliance
Audo AI caps speaker labels at 8 unique speakers per file. For 15-to-30-person town halls, use ElevenLabs Scribe v3 (up to 32 speakers), AssemblyAI (up to 30 speakers on 10+ min files), or Amazon Transcribe (up to 30 speakers). For air-gapped environments with zero outbound API traffic, self-host faster-whisper with pyannote.audio or Speechmatics on-prem containers.
Test Audo AI's transcription API with one keyless curl
Send any audio or video file to POST https://audo.ai/api/try/transcribe with no API key (first minute free, 3x/day), connect https://audo.ai/mcp to your coding agent, or ask for custom volume pricing on /developers.
Questions
What is the cheapest speech-to-text API with speaker diarization in 2026?
On pure pre-recorded list rates as of November 2026, AssemblyAI Universal-2 with standard diarization ($0.15 + $0.02 = $0.17/hr), Gladia Solaria-3 (€0.16/hr across 5 European languages, dropping to €0.12/hr at volume), Google Cloud Speech-to-Text V2 Dynamic Batch ($0.18/hr on Chirp 3 across 14 diarization locales), and ElevenLabs Scribe v3 ($0.20/hr across 90+ languages, up to 32 speakers) are the cheapest raw STT-only APIs. When your workflow also denoises, cuts fillers, bleeps PII, or burns captions onto MP4 video without running your own ffmpeg and GPU stack, Audo AI bundles 8-speaker transcription ($1.80/hr list, $1.13–$1.44/hr on $10–$100 packs) with $0.01/min downstream media chaining in one API.
Why does Audo AI cost $0.03/min ($1.80/hr) when Deepgram and AssemblyAI charge $0.17–$0.31/hr?
Deepgram and AssemblyAI are raw speech-to-text engines that return stateless JSON text and bill multichannel stereo recordings per channel. Audo AI is an end-to-end audio and video processing API: one $0.03/min ($1.13–$1.44/hr on bonus packs, plus custom volume tiers at /developers) run ingests raw video up to 10 GB and 10 hours with no multichannel multiplier, includes 8-speaker diarization plus JSON, TXT, SRT, and WebVTT files, and exposes a reusable transcript_file_id that cuts downstream filler cutting, profanity/PII audio bleeping, highlight finding, chapters, and burned-in captions to $0.01/min.
Is there a free speech-to-text API you can test with curl and no API key?
Yes. Audo AI's Free API lets any developer or agent call POST https://audo.ai/api/try/transcribe with curl and no API key or account to transcribe the first minute of any audio or video file (with speaker labels and JSON, TXT, SRT, or VTT output) 3 times per day. Deepgram ($200), Speechmatics ($100), AssemblyAI ($50), and Gladia (€50) offer generous starter balances without a payment card, but all require account creation and an API key first.
What is happening to OpenAI's Whisper API (whisper-1) and gpt-4o-transcribe-diarize?
OpenAI has deprecated whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize, and /v1/audio/translations with a scheduled shutdown date of February 26, 2027. OpenAI's replacement file model, gpt-transcribe ($0.0045/min or $0.27/hr), retains the 25 MB per-file upload cap and returns plain transcript text plus detected languages only—dropping speaker diarization, word-level timestamps, and SRT/VTT exports—so teams needing those features must migrate to dedicated STT APIs such as Deepgram, AssemblyAI, ElevenLabs Scribe, or Audo AI.
Which transcription APIs charge extra for stereo or multichannel audio?
Deepgram, AssemblyAI, and Google Cloud Speech-to-Text V2 bill multichannel audio by the sum of all channel durations—meaning a 1-hour two-channel stereo recording bills as 2 full hours unless you downmix to mono first. Amazon Transcribe includes up to 2 channels at the single-file duration rate, and Audo AI charges strictly by recording duration with zero multichannel multiplier.
Does Audo AI offer real-time WebSocket streaming or high-volume API discounts?
Audo AI does not offer live WebSocket streaming—it is built for fast file and URL batch processing (under 10 seconds for a 2-minute clip and under 2 minutes for recent 1-hour recordings) alongside REST API and remote MCP workflows. For high-volume batch workloads, developers can request custom lower per-hour rates on /developers#volume and receive an emailed rate within a day with no sales call or minimum contract.