TWELP 400 bps Vocoder - DSP Innovations

Details: Published: 2016 November 09; Last Updated: 2026 September 14

brings 600–700 bps-class speech quality and intelligibility to just 400 bps. Below are comparative test results against MELPe 600 bps and CODEC2 700 bps, which require 50% and 75% more bitrate, respectively.

Deployed in Digital HF Radio, tactical communications, secure voice and other bandwidth-constrained communication applications.

TWELP Technology

Advanced Speech Coding Approach: Pitch-synchronous processing and proprietary Tri-Wave excitation, providing an alternative to MELPe, AMBE+2 and ACELP-based approaches.
Mature & Proven: Nearly two decades of continuous development and real-world deployment in professional communication systems.
Speech & Beyond: High-quality speech plus improved reproduction of tones, sirens, music and other non-speech signals.

Speech Quality

TWELP 400 bps was compared with MELPe at 600 bps and CODEC2 at 700 bps using an updated 8 kHz test set derived from the ITU-T P.50 speech database in 20 languages.

Note:
The original ITU-T P.50 speech files were resampled to 8 kHz. Files for each language were combined and inter-speech pauses were shortened to minimize the effect of silence on the evaluation results.

Speech quality was objectively evaluated using ITU-T P.862 (PESQ):

Across all 20 languages, TWELP 400 bps scores higher than MELPe 600 bps and also achieves a slightly higher average PESQ score than CODEC2 700 bps. Exact PESQ scores are shown below.

Language	TWELP 400 bps	MELPe 600 bps	CODEC2 700 bps
American	2.290	2.184	2.268
Arabic	2.169	2.168	2.213
British	2.317	2.195	2.101
Chinese	2.100	2.054	2.230
Danish	2.256	2.156	2.280
Dutch	2.161	2.069	1.992
Finnish	2.258	2.245	2.027
French	2.330	2.239	2.293
German	2.375	2.225	2.190
Greek	2.242	2.180	2.207
Hindi	2.402	2.334	2.218
Hungarian	2.335	2.259	2.314
Italian	2.450	2.399	2.424
Japanese	2.335	2.251	2.286
Norwegian	2.216	2.205	2.299
Polish	2.300	2.236	2.131
Portuguese	2.428	2.369	2.387
Russian	2.153	2.038	2.029
Spanish	2.394	2.322	2.416
Swedish	2.434	2.420	2.315
Average	2.297	2.227	2.231
The results show that TWELP 400 bps slightly outperforms MELPe 600 bps and CODEC2 700 bps on average while operating at a significantly lower bitrate.

Speech Intelligibility

Speech intelligibility was evaluated using three complementary objective metrics: ASR-based Speech Intelligibility, STOI and ESTOI.

STOI and ESTOI were calculated using the updated 8 kHz test set described above. ASR-based Speech Intelligibility was evaluated separately using the original ITU-T P.50 speech files at their original 16 kHz sampling rate.

ASR reference transcripts were generated by whisper.cpp with the Whisper large-v3 model directly from the original, unprocessed 16 kHz P.50 speech files. For vocoder evaluation, the same speech files were converted to 8 kHz, processed by the vocoder, converted back to 16 kHz and transcribed using the same ASR system.
ASR-based Speech Intelligibility is calculated as 100 - CER (%) by comparing each processed-speech transcript with its corresponding ASR reference transcript.
CER is used instead of WER because it measures character-level recognition errors without treating an entire word as incorrect due to a single-character error.
CER scoring ignores letter case, whitespace and punctuation.

Control measurement: the same 16 kHz → 8 kHz → 16 kHz conversion path, but without any vocoder processing, resulted in an overall ASR-based intelligibility of 99.41%. This shows that narrowband conversion alone produces a small measurable reduction in the ASR-based intelligibility measure, even for clean, unencoded speech.
All vocoder results below are scored against the ASR reference transcripts generated from the original, unprocessed 16 kHz speech, not against transcripts from the converted signal.

Across 20 languages, TWELP 400 bps achieves overall ASR-based speech intelligibility comparable to MELPe 600 bps and clearly higher than CODEC2 700 bps. Exact values are shown below:

Language	TWELP 400 bps	MELPe 600 bps	CODEC2 700 bps
American	94.75	94.13	92.18
Arabic	82.89	71.81	85.23
British	93.38	95.42	92.78
Chinese	84.73	80.30	80.54
Danish	67.41	72.92	64.83
Dutch	89.59	89.98	83.59
Finnish	76.39	77.10	52.36
French	94.85	94.03	93.53
German	94.08	93.37	90.49
Greek	83.31	83.86	73.36
Hindi	77.77	75.98	65.33
Hungarian	88.60	90.88	89.74
Italian	92.57	94.77	89.86
Japanese	93.72	93.87	92.01
Norwegian	91.14	92.66	86.15
Polish	87.95	89.69	84.35
Portuguese	97.04	95.56	91.82
Russian	97.47	95.60	96.39
Spanish	93.43	94.02	89.41
Swedish	68.80	76.79	55.88
Overall	86.79	87.66	81.38
Overall ASR-based speech intelligibility: 86.79% for TWELP 400 bps, versus 87.66% for MELPe 600 bps and 81.38% for CODEC2 700 bps.

Across all 20 languages, TWELP 400 bps achieves higher STOI scores than both MELPe 600 bps and CODEC2 700 bps. Exact values are shown below:

Language	TWELP 400 bps	MELPe 600 bps	CODEC2 700 bps
American	81.15	79.14	75.33
Arabic	78.75	78.42	73.47
British	77.96	75.26	71.45
Chinese	78.87	77.37	76.00
Danish	81.03	79.06	75.35
Dutch	79.49	76.61	71.49
Finnish	75.66	74.75	70.04
French	80.75	78.58	75.55
German	80.93	78.82	74.03
Greek	78.94	77.93	74.01
Hindi	80.04	78.26	72.82
Hungarian	81.09	77.96	75.91
Italian	79.36	78.00	73.17
Japanese	79.72	79.01	76.32
Norwegian	80.64	79.28	77.45
Polish	80.40	78.11	73.52
Portuguese	80.78	77.96	75.42
Russian	78.71	75.33	71.82
Spanish	78.19	77.82	74.64
Swedish	77.48	76.84	72.96
Average	79.50	77.73	74.04
Average STOI scores: 79.50 for TWELP 400 bps, 77.73 for MELPe 600 bps, and 74.04 for CODEC2 700 bps.

ESTOI provides a complementary objective measure of speech intelligibility, extending the STOI analysis with a different treatment of spectro-temporal information:

Across all 20 languages, TWELP 400 bps achieves higher ESTOI scores than both MELPe 600 bps and CODEC2 700 bps. Exact values are shown below:

Language	TWELP 400 bps	MELPe 600 bps	CODEC2 700 bps
American	70.92	69.68	62.85
Arabic	70.23	69.96	61.09
British	68.54	67.07	59.87
Chinese	71.44	69.87	66.90
Danish	71.78	69.72	63.09
Dutch	69.88	69.31	61.75
Finnish	66.96	66.21	59.36
French	71.40	70.05	63.46
German	70.99	68.43	59.97
Greek	71.55	70.61	63.82
Hindi	68.61	67.76	57.26
Hungarian	69.68	67.07	62.83
Italian	69.11	68.54	62.25
Japanese	71.37	70.92	65.92
Norwegian	72.23	70.63	66.61
Polish	71.25	70.38	61.92
Portuguese	71.41	69.39	63.50
Russian	68.66	66.28	60.21
Spanish	70.41	70.33	65.38
Swedish	67.11	67.10	60.06
Average	70.18	68.97	62.41
Average ESTOI scores: 70.18 for TWELP 400 bps, 68.97 for MELPe 600 bps, and 62.41 for CODEC2 700 bps.

The P.862, STOI/ESTOI, and CER scoring utilities, including the CER source code and ASR result analysis scripts, are available in the “Downloads” section at the bottom of this page.
The open-source whisper.cpp ASR engine and the corresponding Whisper large-v3 model can be obtained separately from their official repositories.
The supplied tools and instructions allow the results presented above to be independently reproduced and verified.

Speech Samples (WAV files)
Independent listeners compared TWELP 400 bps with MELPe 600 bps and CODEC2 700 bps in preference listening tests. Listening preferences were closely divided, but TWELP was preferred overall in comparisons with both vocoders, with listeners describing its speech as more natural and clear.

Listen to the source speech and the corresponding CODEC2 700 bps, MELPe 600 bps and TWELP 400 bps samples below.

For more accurate comparison, we recommend using good-quality headphones or speakers.

The complete P.50 sample sets for all languages are also available in the Downloads section at the bottom of this page.

Language	Source speech	CODEC2 700 bps	MELPe 600 bps	TWELP 400 bps
American
Arabic
British
Chinese
Danish
Dutch
Finnish
French
German
Greek
Hindi
Hungarian
Italian
Japanese
Norwegian
Polish
Portuguese
Russian
Spanish
Swedish

Speech & Beyond
Unlike many low-bitrate vocoders optimized primarily for speech, TWELP also provides improved reproduction of non-speech signals, including alert tones, police, ambulance and fire sirens, music and other audio signals.
Combined with natural speech reproduction, this makes TWELP well suited for digital radio and other communication systems where both speech and non-speech signals must be transmitted over a very low-bitrate channel.
Listen to the comparison below:

Source signal	CODEC2 700 bps	MELPe 600 bps	TWELP 400 bps

High Robustness To Acoustic Noise

An additional benefit of the TWELP approach is its strong robustness to acoustic noise. Its pitch-synchronous processing, robust pitch estimation and other features of TWELP technology help preserve speech intelligibility when the input speech is contaminated by background noise.
For this comparison, ESTOI was selected as a single objective intelligibility metric to provide a focused and consistent evaluation of speech under acoustic-noise conditions.

TWELP 400 bps outperforms both MELPe 600 bps and CODEC2 700 bps in speech intelligibility under acoustic noise across every one of the 20 tested languages. Exact ESTOI values are shown below:

Language	TWELP 400 bps	MELPe 600 bps	CODEC2 700 bps
American	63.38	58.30	54.65
Arabic	62.56	59.81	54.86
British	60.10	56.25	51.27
Chinese	65.06	60.16	58.59
Danish	63.01	57.92	53.17
Dutch	60.49	56.16	52.60
Finnish	59.03	55.11	51.52
French	63.99	59.85	55.06
German	61.30	56.66	51.86
Greek	63.43	59.48	55.75
Hindi	59.39	55.65	49.93
Hungarian	63.31	59.02	54.75
Italian	60.28	57.41	53.55
Japanese	65.53	61.41	57.93
Norwegian	65.67	61.08	57.47
Polish	62.44	59.52	53.57
Portuguese	63.10	58.46	55.27
Russian	60.91	56.19	50.70
Spanish	64.21	59.90	57.50
Swedish	56.66	53.45	50.33
Average	62.20	58.09	54.02
Average ESTOI score in acoustic noise: 62.20 for TWELP 400 bps, versus 58.09 for MELPe 600 bps and 54.02 for CODEC2 700 bps.

In addition to this inherent robustness, TWELP includes an NCSE (Noise Cancellation Speech Enhancement) preprocessor that can further reduce input noise and improve speech quality.
The samples below compare heavily noisy English speech (SNR = 10 dB) processed by CODEC2 700 bps, MELPe 600 bps and TWELP 400 bps, first with noise preprocessing disabled and then enabled (MELPe NPP / TWELP NCSE).

NPP NCSE	Input speech (SNR=10dB)	CODEC2 700 bps	MELPe 600 bps	TWELP 400 bps
Disabled
Enabled

The NCSE integrated into the TWELP vocoder is described in more detail on the webpage for our standalone product, 'NCSE-AGC Preprocessor'.

High Robustness To The Channel Errors.

The TWELP technology offers highly efficient speech compression by eliminating redundancy while preserving excellent quality and intelligibility. To enhance robustness against transmission errors, we provide specialized versions called TWELP Robust.
These vocoders are based on an effective Joint Source-Channel Coding approach. Each vocoder is equipped with a custom-designed FEC, tailored to its specific characteristics and operational conditions.
TWELP Robust vocoders provide high speech quality simultaneously in noisy channel as well as in noiseless channel. FEC can operate with "soft decisions" as well as with "hard decisions" from a modem. "Soft decisions" mode provides much better robustness in comparison with the "hard decisions" mode.

For all users of our non-robust vocoder versions, we offer the following recommendations.

The diagram below illustrates the sensitivity of bits at the output of the vocoder to communication channel errors.
Essentially, the diagram shows by what percentage speech quality is reduced when a specific bit is distorted. The first bits in order cause catastrophic distortions, while the latter bits have significantly less impact on quality.

	TWELP 400 bps Vocoder
1	54.87
2	54.31
3	52.83
4	50.99
5	49.94
6	49.93
7	48.49
8	41.27
9	38.88
10	37.63
11	36.97
12	36.15
13	35.26
14	35.26
15	34.53
16	34.40
17	33.37
18	33.18
19	33.14
20	32.06
21	28.72
22	27.14
23	24.47
24	22.99
25	21.69
26	21.56
27	19.69
28	18.33
29	15.11
30	14.32
31	11.93
32	11.50
33	9.29
34	8.98
35	8.33
36	7.85
37	5.43
38	5.07
39	4.31
40	3.31

We strongly recommend using FEC (Forward Error Correction) with unequal protection of the bits in strong accordance with their sensitivity to errors and utilizing 'Soft Decisions' decoding. This will provide the highest robustness of the vocoder against errors in the channel.

Additional Functionalities. The following additional functionalities are developed by DSPINI and integrated into TWELP vocoders:

Noise Cancellation Speech Enhancement (NCSE)
Automatic Gain Control (AGC),
Voice Activity Detector (VAD),
Discontinuous Transmission (DTX),
Tone Detection/Generation (Single tones and Dual tones). The tones are transmitted by the vocoder facilities.

Note:
The Tone Detector/Generator and DTX are not integrated into the code by default but can be added free of charge upon request.

Each functionality has unique features, performance and characteristics, providing significant superiority over any well-known implementations on the market.

Technical Characteristics And Resource Requirements:

Technical characteristics
Bit Rate (bps)	Algorithm	Frame size (ms)	Algorithmic delay (including frame size) (ms)	Sampling rate (kHz)	Signal format	Bit stream format
400	TWELP	100	120	8	Linear 16-bit PCM	40

Additional functionalities
Name	Functionality	Technical characteristics
Name	Functionality	Name	Value
AGC	Automatic Gain Control	Control range:	0 ... +42 dB
NCSE	Noise Canceller - Speech Enhancer	SNR improvement	20 dB
NCSE	Noise Canceller - Speech Enhancer	Speech quality improvement	> 0.1 PESQ
Tone Detector	Single/Dual tones detection	In accordance with international standards
Tone Generator	Single/Dual tones generation	Special generator, kept continuity of signal (phase and amplitude of signal of previous frame)
DTX	Discontinuous Transmission	Reduces bit rate down to 110 bps in pauses between active speech regions
VAD	Voice Activity Detection	High reliability even with pink noise at an SNR < 0 dB.
CNG	Comfort Noise Generation	Spectrally shaped based on the estimated background noise spectrum
CNG	Comfort Noise Generation	Average RMS level	−58.2 dBFS

The NCSE and AGC integrated into the TWELP vocoder are described in more detail on the webpage for our standalone product, 'NCSE-AGC Preprocessor'.

Resources for ARM Cortex-M4 platform
Module	MIPS* peak	Memory (KBytes)
		Program	Data
		Program	Constants	Channel	Heap	Stack
Voice Encoder	49	42	196	4.7	6.7	1.0
NCSE	6.0
AGC	0.5
Voice Decoder	19
Voice Encoder + Voice Decoder	68
Total	74.5

Resources for TI's C64 DSP platform
Module	MIPS* peak	Memory (KBytes)
		Program	Data
		Program	Constants	Channel	Heap	Stack
Voice Encoder	17.6	81	196	4.7	6.7	1.0
NCSE	2.6
AGC	0.3
Voice Decoder	5.3
Voice Encoder + Voice Decoder	22.9
Total	25.8

Resources (estimated) for TI's C55 DSP platform
Module	MIPS* peak	Memory (KBytes)
		Program	Data
		Program	Constants	Channel	Heap	Stack
Voice Encoder	30	26	196	4.7	6.7	1.0
NCSE	6.2
AGC	0.4
Voice Decoder	13
Voice Encoder + Voice Decoder	43
Total	49.6

* DSPINI continues optimization of the TWELP algorithm and code in order to minimize computational complexity of the vocoder.

Software Integrity and Security. DSPINI guarantees the ABSOLUTE integrity of its software, free from any undocumented features, undeclared capabilities, or hidden functions. Our customers can be assured that none of our software/code contains any secret features or functionalities concealed from the user. If necessary, we are ready to provide the source code of our software products for appropriate certification.
Moreover, our software is available in source code form—you simply need to purchase the appropriate license to use it.

Guarantee And Support. DSPINI guarantees a quality and accordance of all technical characteristics of the product to requirement of current specifications. Testing and other method of quality control are used for guarantee support.

Any Platforms. DSPINI can port this vocoder software into any other DSP, RISC or general- purposes platform inshort time: 1-2 months.

Licensing Terms. To use the vocoder, customer should obtain a license from DSPINI only.

Customization. The vocoder can be customized under any specific requirements- other bit rate, frame size, any other robustness to channel errors, etc. Please contact with us for details.

Prospects. DSPINI is impoving and developing continuously a set of new vocoders with range from 300 bps up to 9600 bps, based on TWELP technology.

Related Software. This vocoder may be effectively used in a bundle with other DSPINI's products:

Linear and acoustic echo cancellers,
Multichannel noise cancellers (including two-microphone adaptive array),
Wired or radiomodems for any types of channels and bitrates,
Other products.

Downloads:

Datasheet (pdf)
ITU-T P.50 source speech samples (zip)
TWELP 400 bps speech samples (zip)
MELPe 600 bps speech samples (zip)
CODEC2 700 bps speech samples (zip)
P.862 and STOI/ESTOI utilities
PC-evaluation package (zip) — on request
User's Guide document (pdf) — on request