DeepgramAudio modelActive

Aura-2

PopularityRank 215 of 876

44out of 100same as last month

Summary

Aura-2 is Deepgram’s multilingual text-to-speech family, with voices for English, Spanish, German, French, Dutch, Italian, and Japanese. The model ID used here is aura-2-thalia-en through the /v1/speak endpoint.

With Aura-2, the model identifier also selects a concrete voice. aura-2-thalia-en is the English Thalia variant; German or Japanese output requires another voice identifier. Deepgram also positions Aura-2 for languages that its English-focused Flux TTS does not cover.

GraduallyBench speech output

Automatic text comparison after speech generation.

Each first audio output is fully decoded and transcribed once by gpt-transcribe. Word error rate (WER) and character error rate (CER) compare this transcript with the frozen input script. Lower values mean fewer text differences. Errors can also come from the transcription model.

This model-dependent text proxy does not measure naturalness, voice preference or human intelligibility. Numbers and names depend on the fixed normalization. The requested voice IDs are recorded; the provider does not independently confirm the delivered voice identity.

Protocol audio-tts-asr-content-proxy-v1. Actual provider billing is unknown.

Analysis date Nov 5, 2026

Deepgram Aura 2

Cloudflare Workers AI · @cf/deepgram/aura-2-en

Generated: 16
Fully decoded: 16 / 16
Transcribed once: 16 / 16

ASR text comparison with the input script
Language	Outputs	WER	CER
German	Not tested
English	16 / 16	4.06 %13 / 320	3.09 %49 / 1588

Requested voices asteria · orion

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 189 published benchmark results

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

EmergentTTS-Eval overall WER (Show measurement, test conditions, and source)

Source value: 16.83
Score: 16.83
Metric: Whisper-v3-large word error rate across all six categories
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Lower is better
Sample: 1645
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. Exact paper row: DeepGram Aura-2, voice Thalia-en; pure TTS receives utterance text directly. Whisper-v3-large computes WER across the 1,645-prompt EmergentTTS-Eval dataset.
Context: This is a paper-specific fixed benchmark. The paper's LALM comparisons use Gemini 2.5 Pro; they are model-as-judge results, not human listening scores. The paper reports this model-specific row, not an exact current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval overall win rate vs GPT-4o-mini-TTS Alloy (Show measurement, test conditions, and source)

Source value: 32.44
Score: 32.44
Metric: pairwise model-as-judge preference win rate; ties count as one half
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Higher is better
Sample: 1645
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. Exact paper row: DeepGram Aura-2, voice Thalia-en; pure TTS receives utterance text directly. Candidate audio was compared with gpt-4o-mini-tts using Alloy voice on 1,645 prompts; outputs were randomized and judged by Gemini 2.5 Pro at temperature 0.
Context: Relative to this named baseline, not an absolute quality score. Gemini 2.5 Pro is an automatic audio-language judge, not a human panel. Table 1 reports four Parsing Fail for this model row, but does not state whether failed judgments are excluded from the overall win-rate denominator; the denominator is unresolved. The paper reports a voice row rather than an exact current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval Emotions WER (Show measurement, test conditions, and source)

Source value: 3.45
Score: 3.45
Metric: Whisper-v3-large word error rate for the Emotions category
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Lower is better
Sample: 280
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. DeepGram Aura-2, exact paper voice Thalia-en. Category-specific result from the study's pure-TTS, direct-utterance setup; Whisper-v3-large WER on 280 Emotions prompts.
Context: Fixed study category; the overall evaluation also reports automatic model-as-judge comparisons, not human listening scores. The paper row identifies a voice but not a current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval Emotions win rate vs GPT-4o-mini-TTS Alloy (Show measurement, test conditions, and source)

Source value: 29.28
Score: 29.28
Metric: pairwise model-as-judge preference win rate for the Emotions category; ties count as one half
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Higher is better
Sample: 280
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. DeepGram Aura-2, exact paper voice Thalia-en, compared with gpt-4o-mini-tts Alloy on 280 Emotions prompts. Audio order randomized; Gemini 2.5 Pro judges at temperature 0; ties count as half a win.
Context: Relative to the named GPT-4o-mini-TTS Alloy baseline, not an absolute quality score. Gemini 2.5 Pro is a model judge, not a human panel; the paper row identifies a voice rather than the current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval Foreign Words WER (Show measurement, test conditions, and source)

Source value: 21.41
Score: 21.41
Metric: Whisper-v3-large word error rate for the Foreign Words category
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Lower is better
Sample: 280
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. DeepGram Aura-2, exact paper voice Thalia-en. Category-specific result from the study's pure-TTS, direct-utterance setup; Whisper-v3-large WER on 280 Foreign Words prompts.
Context: Fixed study category containing foreign-language words in English utterances; not a general multilingual-support test. The paper row identifies a voice but not a current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval Foreign Words win rate vs GPT-4o-mini-TTS Alloy (Show measurement, test conditions, and source)

Source value: 18.75
Score: 18.75
Metric: pairwise model-as-judge preference win rate for the Foreign Words category; ties count as one half
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Higher is better
Sample: 280
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. DeepGram Aura-2, exact paper voice Thalia-en, compared with gpt-4o-mini-tts Alloy on 280 Foreign Words prompts. Audio order randomized; Gemini 2.5 Pro judges at temperature 0; ties count as half a win.
Context: Relative to the named GPT-4o-mini-TTS Alloy baseline and only this fixed category. Gemini 2.5 Pro is a model judge, not a human panel; the row identifies a voice rather than the current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval Paralinguistics WER (Show measurement, test conditions, and source)

Source value: 23.73
Score: 23.73
Metric: Whisper-v3-large word error rate for the Paralinguistics category
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Lower is better
Sample: 280
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. DeepGram Aura-2, exact paper voice Thalia-en. Category-specific result from the study's pure-TTS, direct-utterance setup; Whisper-v3-large WER on 280 Paralinguistics prompts.
Context: Fixed study category. The paper's expressiveness comparison uses a model judge and is not human listening; the paper row identifies a voice but not a current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval Paralinguistics win rate vs GPT-4o-mini-TTS Alloy (Show measurement, test conditions, and source)

Source value: 21.14
Score: 21.14
Metric: pairwise model-as-judge preference win rate for the Paralinguistics category; ties count as one half
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Higher is better
Sample: 280
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. DeepGram Aura-2, exact paper voice Thalia-en, compared with gpt-4o-mini-tts Alloy on 280 Paralinguistics prompts. Audio order randomized; Gemini 2.5 Pro judges at temperature 0; ties count as half a win.
Context: Relative to the named GPT-4o-mini-TTS Alloy baseline and only this fixed category. Gemini 2.5 Pro is a model judge, not a human panel; the row identifies a voice rather than the current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval Complex Pronunciation WER (Show measurement, test conditions, and source)

Source value: 54.49
Score: 54.49
Metric: Whisper-v3-large word error rate for the Complex Pronunciation category
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Lower is better
Sample: 245
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. DeepGram Aura-2, exact paper voice Thalia-en. Category-specific result from the study's pure-TTS, direct-utterance setup; Whisper-v3-large WER on 245 Complex Pronunciation prompts.
Context: Fixed, deliberately difficult category for unusual characters, numerals, addresses, equations, abbreviations, and tongue-twisters; not a general-purpose WER estimate. The paper row identifies a voice but not a current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval Complex Pronunciation win rate vs GPT-4o-mini-TTS Alloy (Show measurement, test conditions, and source)

Source value: 33.81
Score: 33.81
Metric: pairwise model-as-judge preference win rate for the Complex Pronunciation category; ties count as one half
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Higher is better
Sample: 245
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. DeepGram Aura-2, exact paper voice Thalia-en, compared with gpt-4o-mini-tts Alloy on 245 Complex Pronunciation prompts. Audio order randomized; Gemini 2.5 Pro judges at temperature 0; ties count as half a win.
Context: Relative to the named GPT-4o-mini-TTS Alloy baseline and only this deliberately difficult category. Gemini 2.5 Pro is a model judge, not a human panel; the paper row identifies a voice rather than the current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval Questions WER (Show measurement, test conditions, and source)

Source value: 1.24
Score: 1.24
Metric: Whisper-v3-large word error rate for the Questions category
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Lower is better
Sample: 280
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. DeepGram Aura-2, exact paper voice Thalia-en. Category-specific result from the study's pure-TTS, direct-utterance setup; Whisper-v3-large WER on 280 Questions prompts.
Context: Fixed study category for interrogative versus declarative prosody; it does not represent broad conversational ability. The paper row identifies a voice but not a current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

EmergentTTS-Eval Questions win rate vs GPT-4o-mini-TTS Alloy (Show measurement, test conditions, and source)

Source value: 48.21
Score: 48.21
Metric: pairwise model-as-judge preference win rate for the Questions category; ties count as one half
Unit: %
Benchmark version: EmergentTTS-Eval (NeurIPS 2025), Table 1
Category: audio
Direction: Higher is better
Sample: 280
Source type: independent-evaluator
Evaluator: EmergentTTS-Eval, NeurIPS 2025
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. DeepGram Aura-2, exact paper voice Thalia-en, compared with gpt-4o-mini-tts Alloy on 280 Questions prompts. Audio order randomized; Gemini 2.5 Pro judges at temperature 0; ties count as half a win.
Context: Relative to the named GPT-4o-mini-TTS Alloy baseline and only this fixed category. Gemini 2.5 Pro is a model judge, not a human panel; the row identifies a voice rather than the current API checkpoint string.

Source EmergentTTS-Eval, NeurIPS 2025

Page 1 of 16

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 83 models

pyannote Community-173

Nemotron-3-Diarization57

Gemini 3.5 Transcribe52

Confucius4-R2T252

NVIDIA Nemotron 3.5 Transcribe52

Granite Speech 5.0 470M TurboCTC52

Kokoro-82M51

Nova-351

Fish Audio S2 Pro51

Granite Speech 5.0 470M TurboCTC NC50

Orukeet49

Lyria 3.547

Gemini 3.5 Live Translate46

Phonon-244

Aura-244

Gemini 3.1 Flash TTS44

MiniMax-Music341

Audio8-ASR-Infinite41

GPT-Realtime 2.141

Maya 2 Native40

pyannote Community-1

Nemotron-3-Diarization

Gemini 3.5 Transcribe

Confucius4-R2T2

NVIDIA Nemotron 3.5 Trans…

Granite Speech 5.0 470M T…

Kokoro-82M

Nova-3

Fish Audio S2 Pro

Granite Speech 5.0 470M T…

Orukeet

Lyria 3.5

Gemini 3.5 Live Translate

Aura-2

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 6, 2026

Aura-2Gemini 3.5 Live Translate 46Phonon-2 44Gemini 3.1 Flash TTS 44MiniMax-Music3 41

Method changes in the timeline, a jump at these points is no market event:

1November 1, 2026: Value only with search or usage data, 30-day attention

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

Aura-2 compared

Similar models

More models of the same kind as Aura-2, chosen by model line, developer, size, price and release date.

All specifications

Architecture

Model class: Text-to-speech; Source: Model class

Input and output

Input: Text; Source: Input
Output: Audio; Source: Output
Languages: Voices for English, Spanish, German, French, Dutch, Italian and Japanese; Source: Languages

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking: Unknown; no source
Watermarking technique: Not known; no source

Licence and access

Access: API; Source: Access

API

API model ID: aura-2-thalia-en; Source: API model ID

Further details

Task: Speech generation; Source: Task
Supported details: The API documentation shows the aura-2-thalia-en voice; Source: Supported details
API endpoint: /v1/speak; the model ID selects a specific voice; Source: API endpoint

Sources and data date

Show all 30 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.

Type	Evidence and data date
Research status	Research date October 15, 2026. 27 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Unresolved	exact-current-checkpoint-in-emergenttts-eval Note EmergentTTS-Eval gives an exact Aura-2 Thalia-en voice row, and current Deepgram docs map aura-2-thalia-en to the profile, but the 2025 paper does not identify a current API checkpoint/version string. Keep the paper result distinct from the current Coval API-voice measurements.
Unresolved	emergenttts-win-rate-denominator Note Table 1 reports four judge Parsing Fail for the DeepGram Aura-2 Thalia-en row, and the paper attributes parsing failures generally to malformed JSON or maximum-token failures. It does not state whether these four are excluded from overall or per-category win-rate denominators, so the percentages are preserved as reported but their valid-judgment denominator is unknown.
Not published	ttsbench-deepgram-serving-revision Original wording The source config identifies aura-2-thalia-en and the Thalia voice, but the benchmark does not bind an immutable Deepgram checkpoint or API serving revision.
Unresolved	exact-serving-snapshot-and-unpublished-method-details Note The captured external source reports a benchmark-specific model or hosted route but does not establish an immutable currently served checkpoint. Unpublished row-level denominators and confidence intervals remain unknown; sample counts were entered only for cells that state them.
Unresolved	API build, prompt identities, percentile method Note The live Coval page specifies rolling 30-day results and sample counts but does not pin the API build or prompts and does not define the percentile estimation method.
Additional source	Deepgram text-to-speech documentation (retrieved September 22, 2026; October 15, 2026) · Deepgram text-to-speech models and languages · Editorial description reviewed October 2, 2026
Additional source	EmergentTTS-Eval, NeurIPS 2025 (retrieved October 15, 2026)
Additional source	Coval Voice AI Benchmarks (retrieved October 15, 2026; October 16, 2026; October 17, 2026; October 19, 2026; October 20, 2026; October 28, 2026; October 29, 2026; October 30, 2026; October 31, 2026; November 1, 2026; November 2, 2026; November 5, 2026; November 6, 2026)
Additional source	Rime Labs TTSBench (retrieved October 17, 2026)1 2
Additional source	Speko independent voice benchmark (retrieved October 17, 2026)
Additional source	Cekura Bench TTS Model Benchmarks (retrieved October 19, 2026)
Additional source	Deepgram (retrieved October 20, 2026)
Additional source	Model metadata (retrieved October 15, 2026)
Additional source	Model metadata (retrieved October 16, 2026)1 2
Additional source	Model metadata (retrieved October 17, 2026)1 2 3
Additional source	Model metadata (retrieved October 20, 2026)
Additional source	Model metadata (retrieved October 28, 2026)1 2 3
Additional source	Model metadata (retrieved November 5, 2026)1 2 3
Additional source	Model metadata (retrieved November 6, 2026)1 2 3
Additional source	OpenRouter: Popularity index (derived), CC BY 4.0
Additional source	Vercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional source	Ollama: cumulative downloads for the popularity index
Additional source	Replicate: cumulative runs for the popularity index
Additional source	Google Trending Now: limited feed selection for Germany and the US
Additional source	Creative Commons Attribution 4.0