REAMDE.md

Speech-to-Text Benchmark Results

Overview

This benchmark evaluates the accuracy of various Speech-to-Text (STT) models on long-form audio transcription. The evaluation is based on a podcast audio file with a professionally transcribed ground truth reference.

Ground Truth: 4,748 words | 24,929 characters

Total Runs Evaluated: 8

Key Findings

Highest Word Accuracy: Local Whisper-Base

• WER: 17.52%
• Word Accuracy: 82.48%
• Provider: Local inference via Buzz
• Model: whisper-base
• Punctuation Score: 21.90%

The local Whisper-Base model achieved the highest word accuracy among tested models while recording the lowest punctuation score.

Highest Punctuation Score: Deepgram Nova-3

• Punctuation Score: 51.17%
• Context Match Accuracy: 32.33%
• Total Punctuation: 698 (ref: 688)
• Word Accuracy: 81.28% (ranked 3rd)

Deepgram Nova-3 recorded the highest punctuation score while maintaining word accuracy within 1.2 percentage points of the top performer.

Detailed Results

Rankings by Word Accuracy

Rank	Provider	Model	WER %	CER %	Word Accuracy %	Punct Score %
1	Local	whisper-base	17.52	5.38	82.48	21.90
2	Local	whisper-base (auto-detect)	17.52	5.38	82.48	21.90
3	Deepgram	nova-3	18.72	7.33	81.28	51.17
4	AssemblyAI	best	18.79	6.24	81.21	48.43
5	OpenAI	whisper-1	19.27	6.40	80.73	44.44
6	Gladia	solaria-1	20.83	6.30	79.17	44.13
7	Speechmatics	slam-1-global-english	21.65	7.15	78.35	38.23
8	Local	whisper-tiny	22.49	8.39	77.51	18.78

Rankings by Punctuation Accuracy

Rank	Provider	Model	Punct Score %	Context Match %	Punct Count	Word Accuracy %
1	Deepgram	nova-3	51.17	32.33	698 / 688	81.28
2	AssemblyAI	best	48.43	33.72	791 / 688	81.21
3	OpenAI	whisper-1	44.44	34.42	911 / 688	80.73
4	Gladia	solaria-1	44.13	22.56	651 / 688	79.17
5	Speechmatics	slam-1	38.23	30.00	1003 / 688	78.35
6	Local	whisper-base	21.90	13.02	292 / 688	82.48
7	Local	whisper-base	21.90	13.02	292 / 688	82.48
8	Local	whisper-tiny	18.78	8.60	288 / 688	77.51

Analysis

Local vs Cloud Performance

Local Models:

• Whisper-Base: 82.48% accuracy (ranked 1st)
• Whisper-Tiny: 77.51% accuracy (ranked 8th)

Cloud Models:

• Highest: Deepgram Nova-3 at 81.28% (ranked 3rd)
• Range: 78.35% - 81.28%

The local Whisper-Base model achieved 1.2 percentage points higher word accuracy than the highest-scoring cloud service.

Language Detection Impact

Run-1 (language specified as "en") and Run-3 (auto-detect) both used whisper-base and achieved identical results (17.52% WER).

Error Type Distribution

Local Whisper-Base (Highest Word Accuracy)

• Hits: 3,960 (83.4%)
• Substitutions: 726 (15.3%)
• Deletions: 62 (1.3%)
• Insertions: 44 (0.9%)

Deepgram Nova-3 (Highest Cloud Word Accuracy)

• Hits: 3,919 (82.5%)
• Substitutions: 615 (13.0%)
• Deletions: 214 (4.5%)
• Insertions: 60 (1.3%)

OpenAI Whisper-1

• Hits: 3,947 (83.1%)
• Substitutions: 695 (14.6%)
• Deletions: 106 (2.2%)
• Insertions: 114 (2.4%)

Character Error Rate (CER) vs Word Error Rate (WER)

All models show CER lower than WER:

• Lowest CER: 5.38% (Local Whisper-Base)
• Highest CER: 8.39% (Local Whisper-Tiny)
• CER/WER ratio: ~0.29 - 0.37

Model Categories

Premium Cloud Services

• Deepgram Nova-3: 81.28% (WER: 18.72%)
• AssemblyAI Best: 81.21% (WER: 18.79%)
• OpenAI Whisper-1: 80.73% (WER: 19.27%)

These three models cluster closely together with less than 1% difference in accuracy.

Specialized Cloud Services

• Gladia Solaria-1: 79.17% (WER: 20.83%)
• Speechmatics SLAM-1: 78.35% (WER: 21.65%)

Local Inference

• Whisper-Base: 82.48% (WER: 17.52%) - Highest word accuracy
• Whisper-Tiny: 77.51% (WER: 22.49%) - Lowest word accuracy

Punctuation Analysis

Local Model Punctuation Performance

Local Whisper models captured 42-43% of punctuation marks (288-292 out of 688):

• Not detected: Exclamation marks, quotation marks, colons
• Periods: 16% detection rate (42 out of 263)
• Commas: 30% detection rate (31 out of 104)

Cloud Services: Punctuation Patterns

Highest Punctuation Score (Deepgram Nova-3):

• Count: 698 vs 688 reference
• 32.33% context match accuracy

Higher Punctuation Counts:

• Speechmatics: 1,003 marks (+315, 46% above reference)
• OpenAI Whisper-1: 911 marks (+223, 32% above reference)
• AssemblyAI: 791 marks (+103, 15% above reference)

Lower Punctuation Count:

• Gladia: 651 marks (-37, 5% below reference)

Conclusions

1. Word accuracy and punctuation tradeoff: Local Whisper-Base achieved highest word accuracy (82.48%) but lowest punctuation score (21.90%).

2. Deepgram Nova-3 performance: Recorded word accuracy of 81.28% (1.2 percentage points below highest) and highest punctuation score (51.17%, 2.3x higher than local models).

3. Cloud vs local punctuation performance: Cloud services scored 38-51% on punctuation compared to 19-22% for local models.

4. Model size impact: Whisper-base achieved 4.97 percentage points higher accuracy than whisper-tiny, with similar punctuation scores.

5. Language detection: Explicit language specification vs auto-detection produced identical results (17.52% WER) for whisper-base on this English audio sample.

Model Selection Considerations

Deepgram Nova-3

• Punctuation score: 51.17% (highest)
• Word accuracy: 81.28%
• Processing time: 3 seconds for 27 minutes
• Punctuation count: 698 vs 688 reference

Local Whisper-Base

• Word accuracy: 82.48% (highest)
• Punctuation score: 21.90% (lowest)
• Zero marginal cost per transcription
• Detected 42% of reference punctuation marks

Local Whisper-Base + Cloud Post-Processing

• Initial transcription: 82.48% word accuracy
• Requires secondary cloud processing for punctuation

Gladia Solaria-1

• Word accuracy: 79.17%
• Punctuation score: 44.13%
• Punctuation count: 651 (5% below reference)

AssemblyAI Best

• Word accuracy: 81.21% (highest among cloud services)
• Punctuation score: 48.43%
• Punctuation count: 791 (15% above reference)

Technical Notes

• Evaluation Metric: Word Error Rate (WER) and Character Error Rate (CER) using jiwer library
• Audio Duration: ~27 minutes (1,637.97 seconds based on Deepgram metadata)
• Reference Quality: Professional human transcription
• Test Type: Single audio file, English language, podcast format