Speech-to-text API comparison

Choose AssemblyAI for the lowest listed example cost: $21 for 100 audio hours.

Updated · 7 sources · Methodology

NeedPickWhy
Lowest example costAssemblyAI$212
Fastest non-streaming processingDeepgramAA measures 623.2 audio seconds processed per second for Nova-3.5
Lowest AA-WER v2ElevenLabsElevenLabs Scribe v2 2.2% / AssemblyAI Universal-3.5 Pro 3.0%5

Performance

ProviderAPI and model IDAA-WER v2 (%)Audio processing speed (×)
AssemblyAI5Universal-3.5 Pro3.0%101.5×
ElevenLabs5Scribe v22.2%72.2×
Deepgram5Nova-35.2%623.2×
OpenAI5GPT Transcribe3.3%45.0×

Documented language coverage

Provider / modelModel-specific coverage
AssemblyAI · Universal-3.5 Pro18 languages; includes English and Spanish2
ElevenLabs · Scribe v290+ languages; includes English and Spanish3
Deepgram · Nova-3The documented multi mode lists 10 languages, including English and Spanish; the English price uses monolingual mode.6
OpenAI · GPT TranscribeEnglish, Spanish and French codes are documented; no fixed model-specific total is published.7

Prices and request cost

ProviderEnglish / audio hourEnglish / 100 hoursMixed English/Spanish / hourMixed English/Spanish / 100 hours
AssemblyAI2$0.21$21$0.21$21
ElevenLabs3$0.22$22$0.22$22
Deepgram · Nova-3 Monolingual1$0.258$25.8$0.312$31.2
OpenAI4$0.27$27$0.27$27
Cost example: 100 hours of English mono audio
AssemblyAI$21
ElevenLabs$22
Deepgram$25.8
OpenAI$27

Options, strengths and limitations

  1. AssemblyAI

    Universal-3.5 Pro on AssemblyAI has an AA-WER/speed reference of 3.0% · 101.5×; the base transcription price is $0.21/audio hour.

    Base price / audio hour
    $0.21
    Example quantity
    100
    Example total
    $21

    Strengths

    • Universal-3.5 Pro supports 18 languages.2

    Limitations

    • Universal-3.5 Pro costs more than Universal-2’s $0.15/hour; the cheaper model has different capability and language coverage.2

    Sources:25

  2. ElevenLabs

    Scribe v2 on ElevenLabs has an AA-WER/speed reference of 2.2% · 72.2×; the base transcription price is $0.22/audio hour.

    Base price / audio hour
    $0.22
    Example quantity
    100
    Example total
    $22

    Strengths

    • Scribe v2 has the lowest AA-WER in this table at 2.2%.5

    Limitations

    • The $0.22/hour Scribe v2 rate excludes keyterm prompting at $0.05/hour and transcript editing at $0.07/hour.3

    Sources:35

  3. Deepgram

    Nova-3 on Deepgram has an AA-WER/speed reference of 5.2% · 623.2×; the base transcription price is $0.258/audio hour.

    Base price / audio hour
    $0.258
    Example quantity
    100
    Example total
    $25.8

    Strengths

    • AA measures 623.2 audio seconds processed per second for Nova-3.5

    Limitations

    • Nova-3’s 5.2% AA-WER is higher than all other models in this table; the non-streaming price excludes streaming and add-ons.15

    Sources:15

  4. OpenAI

    GPT Transcribe on OpenAI has an AA-WER/speed reference of 3.3% · 45.0×; the base transcription price is $0.27/audio hour.

    Base price / audio hour
    $0.27
    Example quantity
    100
    Example total
    $27

    Strengths

    • GPT Transcribe records 3.3% AA-WER, below Nova-3’s 5.2%.5

    Limitations

    • At $0.0045/minute, GPT Transcribe’s duration-based $0.27/hour rate is the highest of these 4 base tariffs.4

    Sources:45

Methodology and sources

The AA-WER v2 and non-streaming speed figures are published AA measurements, not tests run for this article. The main cost chart uses 100 hours of English mono pre-recorded audio, with no add-ons. Deepgram uses Nova-3 Monolingual Pay As You Go at $0.0043/minute. The separate mixed English/Spanish tariff uses Nova-3 Multilingual at $0.0052/minute: $0.312/hour and $31.2 for 100 hours. The other listed base tariffs do not add a separate language surcharge for this bilingual example. GPT Transcribe bills audio duration at $0.0045/minute. WER is lower-is-better; speed is audio seconds processed per second, not streaming response latency. Language counts follow each model’s documentation and are not a language-quality ranking; OpenAI does not publish a fixed total for GPT Transcribe. Cards use ascending English-example cost. Rolling third-party performance values are the batch snapshot identified by the page’s Updated line, not live guarantees. AA speed factors are rolling 7-day medians measured with 10-minute audio.

  1. Deepgram · Nova-3 specifications and pricing
  2. AssemblyAI · Universal-3.5 Pro specifications and pricing
  3. ElevenLabs · Scribe v2 specifications and pricing
  4. OpenAI · GPT Transcribe specifications and pricing
  5. Artificial Analysis · AA-WER v2 non-streaming
  6. Deepgram · Nova-3 language modes
  7. OpenAI · GPT Transcribe language hints

Use Novita LLMs after transcription

Build summaries from transcript text with Novita Model APIs. The speech-to-text endpoints and rates above belong to the named third-party providers.

Start BuildingExplore Models