riverborn/bangla-stt-benchmark-results

Bangla Speech-to-Text Benchmark: Results

Results of an open benchmark of 8 hosted Bangla (Bengali) speech-to-text APIs on 1,001 real audio clips across 13 domains, in batch and real-time streaming mode. Tested in September 2026.

• 📈 Interactive report: riverborn.com/research/bangla-speech-to-text-benchmark
• 🏆 Leaderboard: riverborn/bangla-stt-leaderboard
• 💻 Code, method and full report: github.com/riverbornai/bangla-speech-to-text-benchmark
• 🎧 Audio and reference text: SUST-CSE-Speech/banspeech
• 🏢 Research by: Riverborn Limited, Dhaka, Bangladesh
• 🔖 DOI (all versions): 10.5281/zenodo.22946242

This repository has no audio and no reference transcripts. They belong to BanSpeech. To score a hypothesis again, or to listen to a clip, get it from BanSpeech and join on file_path (see Usage).

The short answer

• Most accurate for recorded audio: Sarvam AI (saarika:v2.5) at 6.9% CER.
• Most accurate for live streaming: Soniox (stt-rt-v5) at 7.4% CER.
• Cheapest: Soniox at $0.10 per audio hour (batch), with the second-best accuracy.
• Fastest live transcript: Sarvam AI. The final text arrives 0.9 s after the speaker stops (median).

Key findings

1. Bangla specialists win. Sarvam (6.9%) and Soniox (8.1%) lead batch, and change places in streaming. The gap between them is small, but statistically real (paired bootstrap).
2. Live can beat batch. The Soniox real-time model (7.4%) was more accurate than its own batch model (8.1%).
3. Big global names trail. OpenAI gpt-4o-transcribe (28.7%) and Whisper large-v3 (30.7%) make 4 to 5 times the errors of the leaders on Bangla.
4. Domain matters a lot. ElevenLabs Scribe is mid-table overall (14.7%), but it gets 47.1% CER on medicine and less than 10% on audiobooks and TV news.
5. Cheapest is near the top. Soniox batch costs $0.10 per audio hour, one tenth of Google Chirp 2, and has the second-best accuracy.
6. Streaming latency varies 5×. After the speaker stops, Sarvam gives the final transcript in 0.9 s (median). OpenAI takes 4.2 s and Gemini Live takes more than 10 s.

Results

Lower is better. CER (character error rate) is the primary metric. Brackets show 95% bootstrap confidence intervals.

Batch (whole-file transcription)

#	Provider	Model	CER	WER	Mean latency	List price / hour
1	Sarvam AI	saarika:v2.5	6.9% [6.2–7.5]	18.2%	0.99 s	$0.31
2	Soniox	stt-async-v5	8.1% [7.3–9.0]	21.5%	7.58 s	$0.10
3	Deepgram	nova-3	10.1% [9.2–11.1]	23.9%	2.15 s	$0.26
4	Google Gemini	gemini-2.5-flash	13.3% [12.2–14.5]	27.0%	3.85 s	at least $0.12
5	ElevenLabs	scribe_v1	14.7% [13.3–16.1]	26.0%	1.39 s	$0.22
6	Google Cloud	chirp_2	21.9% [19.9–24.1]	41.4%	4.13 s	$0.96
7	OpenAI	gpt-4o-transcribe	28.7% [26.4–31.1]	45.1%	1.11 s	$0.36
8	Groq	whisper-large-v3	30.7% [29.2–32.4]	71.7%	0.66 s	$0.11

Streaming (audio sent at real-time pace)

#	Provider	Model	CER	WER	First partial (p50)	Final after speech ends (p50)	List price / hour
1	Soniox	stt-rt-v5	7.4% [6.5–8.3]	19.7%	1.85 s	2.46 s	$0.12
2	Sarvam AI	saaras:v3-realtime	9.0% [7.8–10.3]	20.3%	1.35 s	0.88 s	$0.31
3	Deepgram	nova-3	17.5% [16.3–18.7]	33.3%	2.33 s	1.55 s	$0.46
4	Google Cloud	chirp_2	23.6% [21.5–25.8]	43.8%	7.15 s	2.17 s	$0.96
5	OpenAI	gpt-4o-transcribe (Realtime API)	24.6% [22.1–27.1]	42.4%	3.60 s	4.22 s	$0.36
6	ElevenLabs	scribe_v2_realtime	38.6% [35.9–41.3]	55.4%	3.13 s	1.36 s	$0.39
7	Google Gemini	gemini-2.5-flash-native-audio-preview-12-2025 (Live API)	69.6% [67.7–71.4]	79.2%	n/a	10.44 s	at least $0.35

Latency was measured from Bangladesh and includes network time. Prices are pay-as-you-go list prices in USD, checked on 24 September 2026. The sources for each price are in prices.yaml.

Configs

Config	Rows	Contents
leaderboard_batch	8	CER and WER with 95% CI, coverage, latency, real-time factor, list and effective price
leaderboard_streaming	7	Same as batch, plus time to first partial, time to first final, finalization latency, partial-result stability
by_domain	195	CER for each provider, mode and domain (13 domains, 77 clips each)
batch_predictions	8,006	Each clip and provider: hypothesis text, CER, WER
streaming_predictions	6,996	Each clip and provider: hypothesis text, CER, WER, latency per clip, number of partial and final results
pairwise	13	Paired bootstrap comparison between providers (CER gap with 95% CI and probability that A is better)
clips	1,001	The test set: BanSpeech file_path and domain

The run_id column is the model key in the GitHub code. The provider and model_id columns name the exact API model that we tested.

Usage

Load a leaderboard:

from datasets import load_dataset

batch = load_dataset("riverborn/bangla-stt-benchmark-results", "leaderboard_batch", split="test")
print(batch.to_pandas()[["rank", "provider", "model_id", "cer", "wer"]])

Join the predictions with the BanSpeech references:

import pandas as pd
from datasets import load_dataset

preds = load_dataset("riverborn/bangla-stt-benchmark-results", "batch_predictions", split="test").to_pandas()

# BanSpeech stores each domain as a split. Load only the text columns.
banspeech = load_dataset("SUST-CSE-Speech/banspeech")
refs = pd.concat(
    [banspeech[split].remove_columns("audio").to_pandas() for split in banspeech],
    ignore_index=True,
)

joined = preds.merge(refs[["file_path", "transcription"]], on="file_path", how="left")
print(joined[["provider", "file_path", "transcription", "hypothesis", "cer"]].head())

To transcribe the audio yourself and add a provider, use the scripts in the GitHub repository.

Method

Test set

1,001 clips from BanSpeech (SUST CSE Speech Group), 77 sampled at random from each of its 13 domains. Regional dialect clips are not included. Clips are 16 kHz WAV, 0.7 s to 26.8 s long (median 2.1 s), 50.0 minutes in total, with 7,996 reference words and 43,411 characters. 97% of clips are shorter than 10 seconds, like voice commands and conversation turns.

Domain	Clips	Minutes	Mean length	Words
Audiobooks	77	2.2	1.7 s	409
Biography	77	3.0	2.3 s	396
Celebrity interview	77	5.2	4.1 s	854
Class lecture	77	5.0	3.9 s	943
Documentary	77	3.1	2.4 s	457
Drama series	77	4.4	3.4 s	774
Kids' cartoon	77	3.4	2.7 s	524
Kids' voices	77	6.5	5.0 s	947
Medicine	77	3.5	2.7 s	481
Parliament	77	3.9	3.0 s	596
Political talk show	77	3.0	2.4 s	498
Sports	77	3.3	2.6 s	586
TV news	77	3.4	2.6 s	531
Total	1,001	50.0	3.0 s	7,996

Metrics

• CER (primary): character substitutions, deletions and insertions that change the transcript into the reference, divided by the reference length. WER is the same at word level.
• Corpus-level: total edits across all clips divided by total reference length, not an average of per-clip rates.
• Why CER: Bangla word spacing is not consistent between annotators and engines. One spacing difference counts as two word errors even when every letter is correct, so WER is much higher than CER for every provider.
• Confidence intervals: 10,000 bootstrap resamples of the clips. Comparisons between providers use a paired bootstrap over the clips that both transcribed (pairwise config).
• Coverage: a failed request is not included in the score of that provider. It is not counted as 100% error. Every provider covered at least 99.4% of the clips.

Text normalization

The same normalizer runs on the references and on every transcript before scoring (src/normalize.py): Unicode NFC, remove zero-width characters (ZWNJ, ZWJ, BOM), map Bangla digits ০–৯ to 0–9, replace punctuation (including the danda ।) with a space, lowercase Latin text, then collapse whitespace. The hypothesis column has the raw provider output, before normalization.

Setup

• Batch: one request for each clip, one at a time. Latency is the wall-clock request time.
• Streaming: 16-bit PCM in 100 ms chunks. Each chunk is sent only after that much real time has passed, like a live microphone. Accuracy is scored on the final transcript, so batch and streaming are directly comparable.
• Settings: provider defaults, except for the language. No custom vocabulary and no fine-tuning. Only Gemini batch received a prompt ("Transcribe this Bengali (Bangla) audio verbatim...").
• Environment: a MacBook in Bangladesh on a regular internet connection. Provider servers were in the US, Europe or India. Google Chirp 2 ran in us-central1.

Models and language settings

Provider	Batch model	Streaming model	Language
Sarvam AI	saarika:v2.5	saaras:v3-realtime	bn-IN
Soniox	stt-async-v5	stt-rt-v5	hint bn
Deepgram	nova-3	nova-3	bn
Google Gemini	gemini-2.5-flash	gemini-2.5-flash-native-audio-preview-12-2025 (Live API)	named in prompt (batch), auto-detect (streaming)
ElevenLabs	scribe_v1	scribe_v2_realtime	ben
Google Cloud STT v2	chirp_2	chirp_2	bn-BD
OpenAI	gpt-4o-transcribe	gpt-4o-transcribe (Realtime API)	bn
Groq	whisper-large-v3	none	bn

Pricing notes

• Groq bills at least 10 seconds for each request. On these short clips, it costs 3.4× its list price.
• Sarvam and Google round each request up to the next second. This adds about 17% on these clips.
• Soniox real-time bills the streaming session time, which includes the wait for the final transcript. This adds about 40% on these clips.
• Sarvam prices in rupees, converted at ₹95.96 per US$ (ECB reference rate, 24 September 2026).
• Token-billed prices (Soniox, OpenAI) are estimates. Gemini prices are minimums, because they include only the audio input.

Limitations

• A snapshot of these APIs in September 2026. Providers update their models often. Run the benchmark again before you make a purchase decision.
• Mostly read and broadcast speech. No phone audio, heavy noise, regional dialects or Bangla-English code-switching.
• Language tags are different for each provider (bn, bn-IN, bn-BD, ben). This can have an effect on the results.
• Gemini Live returned no interim transcripts, and its adapter set no language hint.
• ElevenLabs batch requested scribe_v1, which was scheduled for removal. Every request succeeded, so we cannot tell which version served them.
• Latency includes the network distance from Bangladesh to each provider. It will be different from your location.
• Per-domain scores use 77 clips each. Treat small differences in one domain as ties.
• Only hosted APIs were tested. Self-hosted open-weight models are not included in v1.0.

Conflict of interest

Riverborn Limited funded and ran this benchmark, and paid for all API usage. We have no commercial relationship with any provider that we tested: we are not a partner, reseller or investor of any of them. No provider sponsored or reviewed this benchmark before publication. All code, prompts, settings, transcripts and scores are public, so anyone can check the results.

License

This repository is released under CC-BY-NC-4.0, the same license as the BanSpeech source data, because the hypotheses are transcripts of BanSpeech audio. The benchmark code on GitHub is MIT.

Citation

If you use these results, cite the benchmark and the BanSpeech dataset. DOI (all versions): 10.5281/zenodo.22946242.

@software{riverborn2026banglastt,
  title     = {Bangla Speech-to-Text Benchmark: Hosted Bengali ASR APIs on BanSpeech},
  author    = {Uddin, Nasim and Azad, Saiful},
  publisher = {Zenodo},
  year      = {2026},
  doi       = {10.5281/zenodo.22946242},
  url       = {https://doi.org/10.5281/zenodo.22946242}
}

BanSpeech: SUST-CSE-Speech/banspeech, SUST CSE Speech Group, CC-BY-NC-4.0.

About Riverborn

Riverborn Limited is an AI systems company in Dhaka, Bangladesh. We build voice agents, LLM infrastructure and document intelligence. Questions about this benchmark: hello@riverborn.com.