Skip to main content

Fish AudioAudio modelOpen weights

Fish Audio S2 Pro

PopularityTop group about equally popular, 8 models

52out of 100new calculation

based on two data groups, mostly search, range 42 to 62

Summary

Fish Audio S2 Pro is an open speech synthesis model controlled through prosody and emotion tags. It supports multi-speaker dialogue, voice cloning, and broad language coverage, while the released S2 line is presented for API and local-weights use.

Fish Audio cites about 100 milliseconds to first audio. The official S2 sources cite roughly 50 to 80 languages.

GraduallyBench Audio

In the speech output text comparison, the re-transcription of Fish Audio S2 Pro differs from the source by 2.5 % (8 errors over 321 words, German and English). The median of the 20 models with a German and English value is 3.5 %.

Word error rate, German and English

2.5 %

8 errors over 321 reference words

Character error rate

1.6 %

27 errors over 1,685 characters

Scope

16 / 16

valid of planned outputs

Access

OpenRouter

Text comparison, German and English, compared

Differing words after re-transcription in percent (German and English), lower is better

20 of 20 models

Smaller test scope: Fish Audio S2 Pro, Fish Audio S2.1 Pro, MAI-Voice-2, MAI-Voice-2-Flash, Sonic 3.6, Voxtral Mini TTS, GPT-4o Mini TTS.

As of 11/10/2026. Source: GraduallyBench Audio

Fish Audio S2 Pro

Benchmark run
  • The text comparison uses a second speech recognizer as a proxy.
  • Naturalness and voice quality were not evaluated.
  • The voice actually used is not confirmed.
  • The route does not confirm the exact model version.
Show test details

Speech output with text comparison

OpenRouter · German, English

16 valid outputs from 16 attempted and 16 planned test slots.

Speech output text comparison, WER
1.86 %(3 / 161 reference words)German
Speech output text comparison, CER
0.79 %(7 / 891 reference characters)German
Speech output text comparison, WER
3.13 %(5 / 160 reference words)English
Speech output text comparison, CER
2.52 %(20 / 794 reference characters)English
Speech output text comparison, WER
2.49 %(8 / 321 reference words)German, English
Speech output text comparison, CER
1.6 %(27 / 1,685 reference characters)German, English

WER compares words and CER compares characters with the reference. The speech output was re-transcribed with GPT Transcribe. Its exact backend revision is unconfirmed. Read speech and digital silence do not cover spontaneous speech or real background noise. Error rates can exceed 100 %. Where used, the test references were not independently reviewed by humans. Test coverage and methods differ by model. Each metric stands on its own; there is no combined overall score across tests. Data compiled on 11/10/2026. How we test · Download all Decision and Audio results as JSON

All benchmark results

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

Show all 96 published benchmark results

Page 1 of 8

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.

win rate (Show measurement, test conditions, and source)
Source value
81.88
Metric
win rate
Unit
%
Benchmark version
source version not stated
Category
audio
Direction
Higher is better
Source type
vendor-reported
Evaluator
Fish Audio S2 announcement
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Provider listening evaluation
Context
Source-specific observation; it does not establish a compatible cross-model cohort.
Source Fish Audio S2 announcement
Fish Instruction Benchmark task acceptance rate (Show measurement, test conditions, and source)
Source value
93.3
Metric
accuracy
Unit
%
Benchmark version
source version not stated
Category
audio
Direction
Higher is better
Source type
vendor-reported
Evaluator
Fish Audio S2 announcement
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Provider benchmark
Context
Source-specific observation; it does not establish a compatible cross-model cohort.
Source Fish Audio S2 announcement
Soniqo FLEURS English, WER (Show measurement, test conditions, and source)
Source value
3.9
Metric
WER from Qwen3-ASR on generated FLEURS clone audio
Unit
%
Benchmark version
Soniqo FLEURS clone test/en_us, retrieved 2026-10-17
Category
text-to-speech
Direction
Lower is better
Sample
10
Source type
independent-evaluator
Evaluator
Soniqo speech-swift FLEURS clone benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Fish Audio S2 Pro fp16 local voice cloning on FLEURS test/en_us; ten reference/target pairs in English; neutral plain text, 16 kHz reference and 44.1 kHz generated output. Qwen3-ASR scores WER/CER, utmos22_strong predicts UTMOS, and Soniqo speech embed-speaker with MLX calculates speaker cosine. The current conversion card maps to fishaudio/s2-pro but does not prove the evaluated weight revision.
Context
Soniqo publishes the evaluation and also maintains its speech stack. This is a small local FP16 MLX conversion test, not Fish Audio's hosted API. The current Hugging Face conversion card names fishaudio/s2-pro as its base and resolves to revision 254339851f9f39b68d06cc9284addab5fb32be13, but the report does not bind the tested weights to that revision. UTMOS is predicted, not human MOS; Qwen3-ASR WER/CER and Soniqo speaker embeddings are proxy measures. No confidence interval is published.
Source Soniqo speech-swift FLEURS clone benchmark
Soniqo FLEURS German, WER (Show measurement, test conditions, and source)
Source value
7
Metric
WER from Qwen3-ASR on generated FLEURS clone audio
Unit
%
Benchmark version
Soniqo FLEURS clone test/de_de, retrieved 2026-10-17
Category
text-to-speech
Direction
Lower is better
Sample
10
Source type
independent-evaluator
Evaluator
Soniqo speech-swift FLEURS clone benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Fish Audio S2 Pro fp16 local voice cloning on FLEURS test/de_de; ten reference/target pairs in German; neutral plain text, 16 kHz reference and 44.1 kHz generated output. Qwen3-ASR scores WER/CER, utmos22_strong predicts UTMOS, and Soniqo speech embed-speaker with MLX calculates speaker cosine. The current conversion card maps to fishaudio/s2-pro but does not prove the evaluated weight revision.
Context
Soniqo publishes the evaluation and also maintains its speech stack. This is a small local FP16 MLX conversion test, not Fish Audio's hosted API. The current Hugging Face conversion card names fishaudio/s2-pro as its base and resolves to revision 254339851f9f39b68d06cc9284addab5fb32be13, but the report does not bind the tested weights to that revision. UTMOS is predicted, not human MOS; Qwen3-ASR WER/CER and Soniqo speaker embeddings are proxy measures. No confidence interval is published.
Source Soniqo speech-swift FLEURS clone benchmark
Soniqo FLEURS Modern Standard Arabic, WER (Show measurement, test conditions, and source)
Source value
26.7
Metric
WER from Qwen3-ASR on generated FLEURS clone audio
Unit
%
Benchmark version
Soniqo FLEURS clone test/ar_eg, retrieved 2026-10-17
Category
text-to-speech
Direction
Lower is better
Sample
10
Source type
independent-evaluator
Evaluator
Soniqo speech-swift FLEURS clone benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Fish Audio S2 Pro fp16 local voice cloning on FLEURS test/ar_eg; ten reference/target pairs in Modern Standard Arabic; neutral plain text, 16 kHz reference and 44.1 kHz generated output. Qwen3-ASR scores WER/CER, utmos22_strong predicts UTMOS, and Soniqo speech embed-speaker with MLX calculates speaker cosine. The current conversion card maps to fishaudio/s2-pro but does not prove the evaluated weight revision.
Context
Soniqo publishes the evaluation and also maintains its speech stack. This is a small local FP16 MLX conversion test, not Fish Audio's hosted API. The current Hugging Face conversion card names fishaudio/s2-pro as its base and resolves to revision 254339851f9f39b68d06cc9284addab5fb32be13, but the report does not bind the tested weights to that revision. UTMOS is predicted, not human MOS; Qwen3-ASR WER/CER and Soniqo speaker embeddings are proxy measures. No confidence interval is published.
Source Soniqo speech-swift FLEURS clone benchmark
Soniqo FLEURS Spanish, WER (Show measurement, test conditions, and source)
Source value
6.8
Metric
WER from Qwen3-ASR on generated FLEURS clone audio
Unit
%
Benchmark version
Soniqo FLEURS clone test/es_419, retrieved 2026-10-17
Category
text-to-speech
Direction
Lower is better
Sample
10
Source type
independent-evaluator
Evaluator
Soniqo speech-swift FLEURS clone benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Fish Audio S2 Pro fp16 local voice cloning on FLEURS test/es_419; ten reference/target pairs in Spanish; neutral plain text, 16 kHz reference and 44.1 kHz generated output. Qwen3-ASR scores WER/CER, utmos22_strong predicts UTMOS, and Soniqo speech embed-speaker with MLX calculates speaker cosine. The current conversion card maps to fishaudio/s2-pro but does not prove the evaluated weight revision.
Context
Soniqo publishes the evaluation and also maintains its speech stack. This is a small local FP16 MLX conversion test, not Fish Audio's hosted API. The current Hugging Face conversion card names fishaudio/s2-pro as its base and resolves to revision 254339851f9f39b68d06cc9284addab5fb32be13, but the report does not bind the tested weights to that revision. UTMOS is predicted, not human MOS; Qwen3-ASR WER/CER and Soniqo speaker embeddings are proxy measures. No confidence interval is published.
Source Soniqo speech-swift FLEURS clone benchmark
Soniqo FLEURS Simplified Chinese, CER (Show measurement, test conditions, and source)
Source value
6.2
Metric
CER from Qwen3-ASR on generated FLEURS clone audio
Unit
%
Benchmark version
Soniqo FLEURS clone test/cmn_hans_cn, retrieved 2026-10-17
Category
text-to-speech
Direction
Lower is better
Sample
10
Source type
independent-evaluator
Evaluator
Soniqo speech-swift FLEURS clone benchmark
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Fish Audio S2 Pro fp16 local voice cloning on FLEURS test/cmn_hans_cn; ten reference/target pairs in Simplified Chinese; neutral plain text, 16 kHz reference and 44.1 kHz generated output. Qwen3-ASR scores WER/CER, utmos22_strong predicts UTMOS, and Soniqo speech embed-speaker with MLX calculates speaker cosine. The current conversion card maps to fishaudio/s2-pro but does not prove the evaluated weight revision.
Context
Soniqo publishes the evaluation and also maintains its speech stack. This is a small local FP16 MLX conversion test, not Fish Audio's hosted API. The current Hugging Face conversion card names fishaudio/s2-pro as its base and resolves to revision 254339851f9f39b68d06cc9284addab5fb32be13, but the report does not bind the tested weights to that revision. UTMOS is predicted, not human MOS; Qwen3-ASR WER/CER and Soniqo speaker embeddings are proxy measures. No confidence interval is published.
Source Soniqo speech-swift FLEURS clone benchmark
Fish Audio blind TTS provider comparison / aggregate win rate (Show measurement, test conditions, and source)
Source value
65.7
Metric
Overall share of preference comparisons won
Unit
%
Benchmark version
Fish Audio production-traffic blind comparison, 2026-03-26 to 2026-04-05
Category
text-to-speech preference
Direction
Higher is better
Source type
vendor-reported
Evaluator
Fish Audio, We Blind-Tested Our TTS Against Every Major Competitor
Status
active
Retrieved at
2026-10-17
Methodology
Vendor-reported result. Exact Fish Audio source row: S2 Pro (latest), route fish:s2-pro. Fish Audio ran a ten-day blind A/B test on real web-platform production traffic from March 26 through April 5, 2026. The same text and reference voice were presented; the behavioral winner was the only clip downloaded after the user played both at least twice. The aggregate results table reports the overall win rate. The source does not publish an immutable model revision or a per-cell pair denominator.
Context
Vendor-reported evaluation run by Fish Audio, not an independent model test. API users were excluded; approximately 30% of participants were returning Fish Audio users. The source reports 4,573 in a Samples column but does not define it as a pair denominator, so N is omitted. No confidence interval is published for this cell.
Source Fish Audio, We Blind-Tested Our TTS Against Every Major Competitor
Fish Audio blind TTS provider comparison / win rate vs ElevenLabs V3 (Show measurement, test conditions, and source)
Source value
60
Metric
Pairwise share of preference comparisons won by Fish Audio S2 Pro
Unit
%
Benchmark version
Fish Audio production-traffic blind comparison, 2026-03-26 to 2026-04-05
Category
text-to-speech preference
Direction
Higher is better
Sample
581
Source type
vendor-reported
Evaluator
Fish Audio, We Blind-Tested Our TTS Against Every Major Competitor
Status
active
Retrieved at
2026-10-17
Methodology
Vendor-reported result. Exact Fish Audio source row: S2 Pro (latest), route fish:s2-pro, compared with ElevenLabs V3 in the named head-to-head. Fish Audio ran a ten-day blind A/B test on real web-platform production traffic from March 26 through April 5, 2026. The same text and reference voice were presented; the behavioral winner was the only clip downloaded after the user played both at least twice. This cell covers 581 preference pairs. The source does not publish an immutable Fish model revision.
Context
Vendor-reported evaluation run by Fish Audio, not an independent model test. API users were excluded; approximately 30% of participants were returning Fish Audio users. No confidence interval is published for this cell.
Source Fish Audio, We Blind-Tested Our TTS Against Every Major Competitor
Fish Audio blind TTS provider comparison / win rate vs Inworld TTS 1.5 Max (Show measurement, test conditions, and source)
Source value
80
Metric
Pairwise share of preference comparisons won by Fish Audio S2 Pro
Unit
%
Benchmark version
Fish Audio production-traffic blind comparison, 2026-03-26 to 2026-04-05
Category
text-to-speech preference
Direction
Higher is better
Sample
261
Source type
vendor-reported
Evaluator
Fish Audio, We Blind-Tested Our TTS Against Every Major Competitor
Status
active
Retrieved at
2026-10-17
Methodology
Vendor-reported result. Exact Fish Audio source row: S2 Pro (latest), route fish:s2-pro, compared with Inworld TTS 1.5 Max in the named head-to-head. Fish Audio ran a ten-day blind A/B test on real web-platform production traffic from March 26 through April 5, 2026. The same text and reference voice were presented; the behavioral winner was the only clip downloaded after the user played both at least twice. This cell covers 261 preference pairs. Inworld was restricted to plain text and MP3 output, and only eligible cloned voices could be routed. The source does not publish an immutable Fish model revision.
Context
Vendor-reported evaluation run by Fish Audio, not an independent model test. API users were excluded; approximately 30% of participants were returning Fish Audio users. The source says Inworld received a smaller, plain-text/MP3-only eligible subset because of route constraints, so this is not an even-coverage model ranking. No confidence interval is published for this cell.
Source Fish Audio, We Blind-Tested Our TTS Against Every Major Competitor
Fish Audio blind TTS provider comparison / win rate vs MiniMax Speech 2.8 HD (Show measurement, test conditions, and source)
Source value
95
Metric
Pairwise share of preference comparisons won by Fish Audio S2 Pro
Unit
%
Benchmark version
Fish Audio production-traffic blind comparison, 2026-03-26 to 2026-04-05
Category
text-to-speech preference
Direction
Higher is better
Sample
142
Source type
vendor-reported
Evaluator
Fish Audio, We Blind-Tested Our TTS Against Every Major Competitor
Status
disputed
Retrieved at
2026-10-17
Methodology
Vendor-reported result. Exact Fish Audio source row: S2 Pro (latest), route fish:s2-pro, compared with MiniMax Speech 2.8 HD in the named head-to-head. Fish Audio ran a ten-day blind A/B test on real web-platform production traffic from March 26 through April 5, 2026. The same text and reference voice were presented; the behavioral winner was the only clip downloaded after the user played both at least twice. This cell covers 142 preference pairs. The source does not publish an immutable Fish model revision.
Context
Disputed vendor-reported comparison. Fish Audio says it suspects the MiniMax API integration may not have been optimal, widened routing eligibility mid-test, and early-stopped the experiment; this score describes that tested route/setup and should not be read as an uncontested model-quality ranking. API users were excluded, approximately 30% of participants were returning Fish Audio users, and no confidence interval is published for this cell.
Source Fish Audio, We Blind-Tested Our TTS Against Every Major Competitor
Seed-TTS Eval WER, Chinese (Show measurement, test conditions, and source)
Source value
0.54
Metric
word error rate
Unit
%
Benchmark version
source version not stated
Category
audio
Direction
Lower is better
Source type
vendor-reported
Evaluator
Fish Audio S2 announcement
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Provider S2 evaluation
Context
Source-specific observation; it does not establish a compatible cross-model cohort.
Source Fish Audio S2 announcement

Page 1 of 8

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 54 models

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

Over time

Weekly values from October 18, 2026 to November 20, 2026

Fish Audio S2 ProGemini 3.5 Transcribe 52Nova-3 51Phonon-2 51MiniMax-Music3 50

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

  1. 1November 1, 2026: Value only with search or usage data, 30-day attention
  2. 2November 1, 2026: Method version 4 (common scale, shrinkage)
  3. 3November 9, 2026: New topics measurement V20 (daily quorum, short forms)

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

All specifications

Architecture

Model class
Speech synthesis
Source: Model class
Checkpoint
fishaudio/s2-pro
Source: Checkpoint

Input and output

Input
Text with prosody and emotion tags
Source: Input
Output
Multilingual speech
Source: Output
Languages
About 50 to 80 languages depending on the documentation variant
Source: Languages
Latency
About 100 milliseconds to first audio
Source: Latency

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 and local weights
Source: Access
Shutdown
None announced
no source

Further details

Features
Inline prosody, emotion, multi-speaker dialogue, and voice cloning
Source: Features

Sources and data date

Show all 31 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.

TypeEvidence and data date
Research statusResearch date October 4, 2026. 24 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published

per-model-denominator-and-confidence-interval

Note

The current Hume TTS page does not publish a per-model sample count or confidence interval for the frontier score.

Not published

Controlled Voice serving revision and sample semantics

Note

Artificial Analysis does not pin the serving revision or define the Samples counting unit for this controlled-voice board.

Not published

Serving revision and sample semantics

Note

The leaderboard does not pin the API/model serving revision, and its Samples column does not define the count unit.

Not published

Voice Replication per-model N and uncertainty

Note

Hume does not publish per-model sample counts or confidence intervals for these table cells.

Additional sourceFish Audio S2 open-source announcement (retrieved October 3, 2026; October 4, 2026; October 17, 2026) · Fish Audio S2 open-source announcement · Editorial description reviewed October 3, 2026
Additional sourceFish Audio S2.1 API announcement (retrieved October 3, 2026) · Fish Audio S2.1 API announcement · Editorial description reviewed October 3, 2026
Additional sourceHume AI Real World VoiceEQ (retrieved October 17, 2026)
Additional sourceArtificial Analysis Controlled Voice Arena (retrieved October 17, 2026; October 28, 2026; October 29, 2026; November 3, 2026)
Additional sourceHume Voice Replication Leaderboard (retrieved October 17, 2026)
Additional sourceVoiceArena TTS Leaderboard (retrieved October 17, 2026)
Additional sourceSoniqo speech-swift FLEURS clone benchmark (retrieved October 17, 2026)
Additional sourceFish Audio, We Blind-Tested Our TTS Against Every Major Competitor (retrieved October 17, 2026)
Additional sourceArtificial Analysis Provider Voice Arena leaderboard (retrieved October 19, 2026)
Additional sourceOpenBMB UltraEval-Audio (retrieved October 20, 2026)
Additional sourceFish Audio S2 Technical Report (retrieved October 23, 2026)
Additional sourceArtificial Analysis (retrieved October 28, 2026; October 29, 2026; November 3, 2026)
Additional sourceVoiceArena primary public leaderboard (retrieved October 28, 2026)
Additional sourceModel metadata (retrieved October 17, 2026)1 (unavailable on November 2, 2026) https://voicearena.com/assets/Methodology-CJGHT0Ur.js234
Additional sourceModel metadata (retrieved October 28, 2026)12345678
Additional sourceOpenRouter: Popularity index (derived), CC BY 4.0
Additional sourceVercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional sourceOllama: cumulative downloads for the popularity index
Additional sourceReplicate: cumulative runs for the popularity index
Additional sourceGoogle Trending Now: limited feed selection for Germany and the US
Additional sourceCreative Commons Attribution 4.0